The cold substellar frontier in JWST Extragalactic surveys: a robust parallax measurement yields the most distant known ultra-cold Y dwarf

Introduction

Trigonometric parallax remains the only model-independent route to derive the distances of individual brown dwarfs and isolated planetary-mass objects, and such distance measurements are of crucial importance for inferring their physical properties. As such objects never reach a stable hydrogen-burning main sequence and instead cool continuously, their physical properties at any time are governed by a combination of mass, age, and composition (Burrows et al. 2001; Bowler 2016; Dupuy & Liu 2017; Chabrier et al. 2023). Measurements of their distances allow us to convert their observed apparent magnitudes into absolute magnitudes and ultimately luminosities. This provides an anchor for empirical spectral type–magnitude and colour–magnitude relations, which can then be used to deduce the physical properties of objects without parallaxes

(e.g., [22] (Dupuy & Liu 2012; Dupuy & Kraus 2013; Faherty et al. 2012; Fontanive et al. 2023; Fontanive et al. 2025; Fontanive et al. 2026)). Substellar objects with well-constrained properties supply the valuable empirical benchmarks against which ultracool atmosphere and evolutionary models are tested (e.g. Phillips et al. 2020; Marley et al. 2021; Lacy & Burrows 2023; Beiler et al. 2024b; Mukherjee et al. 2024; Leggett et al. 2025; Mang et al. 2026; Radcliffe et al. 2026), and volume-complete samples further allow us to place crucial constraints on the substellar luminosity function, potentially providing further insight into the low-mass cut-off of the stellar initial mass function (IMF) (e.g. [23] (Kirkpatrick et al. 2012; Kirkpatrick et al. 2021; Hainline et al. 2026b)). Accurate distances also allow measurements of proper motion to be converted into tangential velocities, and hence can provide valuable kinematic information (e.g. Faherty et al. 2009; Burningham et al. 2013; Fontanive et al. 2025).

These measurements have, however, so far been confined to the immediate solar neighbourhood for the coldest brown dwarfs, i.e., late-T and Y dwarfs. They are intrinsically fainter and predominantly radiate into the thermal near and mid-infrared (IR). Y dwarfs, with $T_{\mathrm{eff}} \lesssim500$ K and $\log(L_{\mathrm{bol}}/L_{\odot}) \lesssim-6$ (e.g. (Dupuy & Kraus 2013; Leggett et al. 2021; Fontanive et al. 2026)), emit the bulk of their flux at 4–5 $\mu$m, where the thermal background renders ground-based astrometry difficult and where the WISE survey that discovered them is complete only to $\sim$10–15 pc ((Cushing et al. 2011; Kirkpatrick et al. 2012, 2019)). Measurements of Y dwarf parallaxes have consequently required dedicated space-based campaigns with Spitzer and HST (e.g. (Marsh et al. 2013; Dupuy & Kraus 2013; Beichman et al. 2014; Martin et al. 2018; Kirkpatrick et al. 2019, 2021; Fontanive et al. 2021, 2025; Marocco et al. 2026)). The result of their faintness and the consequent difficulty to measure parallaxes is that only $\sim$50 Y dwarfs are known to date, all within 20 pc of the Sun, and even fewer have a measured parallax.

Within the Y-class, the census of known objects is even less established towards later subtypes. Brown dwarfs of spectral types later than Y1 are known only in small numbers, essentially all within $\sim$15 pc ((Kirkpatrick et al. 2021; Leggett 2026)), and the cold extreme of the sequence almost solely relies on a single object, WISE J085510.83−071442.5 at 2.28 pc (WISE 0855 for short; (Luhman 2014; Skemer et al. 2016; Kirkpatrick et al. 2019)). The absolute magnitude, luminosity, and effective temperature scales for these subtypes are consequently limited, calibrated by only a handful of parallax measurements ((Beiler et al. 2024b; Fontanive et al. 2025)). This impacts our interpretation of their atmospheres, which have effective temperatures overlapping those of the Solar system giant planets and the coldest directly imaged exoplanets (e.g., [12] (Matthews et al. 2024; Lagrange et al. 2025; Crotts et al. 2025)), and are thus expected to have atmospheres with complex molecular absorption features, disequilibrium chemistry, variability, and even water clouds in the lowest-temperature cases (e.g. [128] (Morley et al. 2014; Miles et al. 2020; Phillips et al. 2020; Mang et al. 2022; Lacy & Burrows 2023; Tan et al. 2025; Fontanive et al. 2026; Leggett 2026)). Expanding the number of parallax-measured extremely cold brown dwarfs, particularly those comparable to WISE 0855, is therefore a key goal of substellar science. For Galactic studies, these objects trace the low-mass terminus of the star-formation process, where the shape of the mass function encodes the physics of fragmentation and the minimum mass for opacity-limited collapse ((Kirkpatrick et al. 2019, 2024)). A parallax-calibrated sample beyond the local volume would also provide not only a larger census, but also a direct probe of a different Galactic environment ((Burgasser et al. 2024; Hainline et al. 2026b)).

Deep extragalactic imaging with JWST has opened an unanticipated route to these objects. Surveys such as JADES ((Eisenstein et al. 2026)), CEERS ((Finkelstein et al. 2023)), PRIMER ((Dunlop et al., in prep.)), COSMOS-Web ((Casey et al. 2023)), and UNCOVER ((Benza-son et al. 2024)) reach of order nJy 1–$\sigma$ depths in 1–5 $\mu$m bands in which cold brown-dwarf spectral energy distributions (SEDs) peak, and their fields are revisited repeatedly across cycles (e.g., follow-up of PRIMER with MINERVA; (Muzzin et al. 2025)). At the high Galactic latitudes probed by these surveys, many cold substellar objects have now been detected: photometric searches have produced dozens of T- and Y-type candidates at implied distances of $\sim$0.1–4 kpc ((Nonino et al. 2023; Wang et al. 2023; Hainline et al. 2024a, 2026b; Chen et al. 2025a)), a growing subset of which have been confirmed spectroscopically, several revealing subsolar metallicities consistent with thick-disc or halo membership ((Langeroodi & Hjorth 2023; Burgasser et al. 2024; Hainline et al. 2024b; Tu et al. 2025b,a; Li et al. 2026; Morrissey et al. 2026)). Indeed, cold dwarfs are effective interlopers in extragalactic samples. Their red near-IR colours mimic those of high-redshift galaxies and dust-reddened active galactic nuclei (AGN), a degeneracy familiar from HST deep-field work ((Pirzkal et al. 2005; Dunlop 2013; Aganze et al. 2022)) and now recurring in JWST samples. Photometrically selected candidate “little red dots” (LRDs), massive (stellar mass $M_{*} > 10^{10.5} M_{\odot}$) $z > 4$ galaxies and ultra-high-redshift ($z > 20$) galaxies are all classes of object that have recently proven to be substellar objects (e.g., (Langeroodi & Hjorth 2023; Hainline et al. 2024b; Tu et al. 2025b)). An outstanding case is that of Capotauro, a redshift $z \sim32$ candidate ((Gandolfi et al. 2026b)) in the JWST Cosmic Evolution Early Release Science Survey (CEERS) field, which has been recently confirmed to be a cold brown dwarf through astrometric proper motion detection by Liu et al. (2026a). From an extragalactic perspective, the aforementioned deficiency of a large, parallax-calibrated late-T and Y-dwarf templates for SED fitting therefore compounds the misinterpretation.

The multi-epoch cadence of these fields has already yielded some of the most precise astrometry achieved for cold substellar objects. Proper motions have been measured for a growing number of deep-field brown dwarfs, from HST–JWST and JWST–JWST baselines, with implied transverse velocities of tens to a couple of hundred km s$^{-1}$ ((Hainline et al. 2024a,b, 2026b; Liu et al. 2026a)). A proper motion alone, however, leaves the distance, and therefore the luminosity and tangential velocity, resting on model-dependent spectrophotometric estimates, which for cold brown dwarfs are uncertain at the tens-of-percent level and degenerate with metallicity and gravity ((Hainline et al. 2024a; Beiler et al. 2024b; Mukherjee et al. 2024; Leggett et al. 2025)). Parallax measurements for objects at such distances require milliarcsecond-level relative astrometry with adequate sampling of the parallactic ellipse, which mosaic-level catalogue positions and sparse annual revisits have not so far delivered. To date, no parallax has been reported for any brown dwarf discovered in a JWST deep extragalactic field.

In this paper, we report such a measurement: the first robust parallax detection for the ultracool, distant brown dwarf JADES-GS-BD-5 in the JWST Advanced Deep Extragalactic Survey (JADES) Great Observatories Origins Deep Survey (GOODS)-South field. JADES-GS-BD-5, hereafter GSBD5, was initially reported by Hainline et al. (2024a) at ICRS J2000 coordinates RA = 03:32:20.170 and Dec = −27:50:21.65, and was identified photometrically as a cold brown-dwarf candidate with an overall motion of 50 $\pm$ 20 mas yr$^{-1}$. It was indeed the coldest dwarf found in the JADES fields (Hainline et al. 2026b), and recent spectroscopic follow-up classified it as a Y0–Y1 dwarf (Hainline et al. 2026a). Here, we use the latest available JWST imaging and spectroscopic data to measure its astrometric motion and constrain its physical properties. We further analyse a broader JADES Y-dwarf sample, and demonstrate the current frontier of the ultracool Galactic substellar population in deep JWST extragalactic survey fields.

This paper is structured as follows. In § 2 we describe the NIRCam, MIRI and NIRSpec data and their reduction. We present the astrometric measurement in § 3, including the relative and absolute reference-frame calibrations and the joint proper-motion and parallax fit. The photometric measurement of GSBD5 with both NIRCam and MIRI is then presented in § 4. In § 5, we investigate the spectral type of GSBD5, estimate its physical properties with evolutionary models, and explore atmospheric models to reproduce its spectrophotometric SED. In § 6, we discuss the evidence of GSBD5 being an isolated planetary mass object and an ultracool giant-planet analogue, the possible physical scenarios needed to explain its unique SED, and a sample of nine other Y dwarf candidates discovered in JADES in the context of using JWST extragalactic survey fields to survey this coldest dwarf population.

We summarise our conclusions in § 7. All magnitudes are quoted in the AB system (Oke 1974; Oke & Gunn 1983) unless stated otherwise.

Data

Imaging data

Here we briefly describe the imaging data targeting the sample of brown dwarf candidates in the GOODS-South and GOODS-North areas, including our main target, GSBD5, and other similar objects discussed in later sections. We use our own reductions for all NIRCam imaging data sets, using the software PENCIL, which is a customised version of JWST pipeline version 1.18.0 (Magee et al., in prep). We use the Calibration Reference Data Syste (CRDS) context jwst_1414.pmap. In Table 1, we provide a summary of the different imaging programs and observation dates for each source covered by our study. Pixels are approximately $0.031^{\prime\prime} \times0.031^{\prime\prime}$ in NIRCam short-wavelength (SW) bands, i.e., F070W to F210M; $0.063^{\prime\prime} \times0.063^{\prime\prime}$ in NIRCam long-wavelength (LW) bands, i.e., F250M to F444W; and $0.060^{\prime\prime} \times0.060^{\prime\prime}$ in MIRI bands.

Source ID	PID	Observation Dates
JADES-GS-BD-4	1180	Oct 2022
	1286	Oct 2023
	4540	Oct 2024
JADES-GS-BD-5	1180	Oct 2022
	1210	Oct 2022
	1895	Nov 2022
	2198	Nov 2022
	2514	Oct 2023
	3215	Oct 2023
	4540	Oct 2024
JADES-GS-BD-8	1180	Oct 2022, Oct 2023, Nov 2023, Jan 2024
	1895	Nov 2022
	6541	Nov 2023
	6630	Oct 2024
	8882	Oct 2025, Dec 2025
JADES-GS-BD-18	2514	Oct 2023
JADES-GS-BD-21	1180	Oct 2022, Oct 2023
	1286	Oct 2023
	2514	Oct 2023
JADES-GN-BD-5	1181	Feb 2023
	1895	Feb 2023
	2674	Mar 2023
	3577	Feb 2024
	7404	Jan 2026, Mar 2026, Apr 2026
JADES-GN-BD-8	1181	May 2023
	1264	Dec 2023
	5398	Feb 2025, May 2025
	7404	Jan 2026
JADES-GN-BD-9	1181	Feb 2023
	2674	Mar 2023
	6434	Jan 2025
JADES-GN-BD-11	1181	Feb 2023
	2514	Mar 2023
	6434	Jan 2025, Feb 2025
JADES-GN-BD-16	1181	Apr 2023
	2514	Mar 2023
	2674	Mar 2023

NIRCam and MIRI imaging for GSBD5

There have been seven observing programs with the JWST (Gardner et al. 2006, 2023) Near Infrared Camera (NIRCam, Rieke et al. 2003, 2023a), spanning a $\simeq2$ year window that have targeted the region within which GSBD5 resides. We group them as five epochs.

In Epoch 1, as part of the JADES survey (Rieke et al. 2023b; Eisenstein et al. 2026) GTO 1180 (PI Eisenstein), ultra-deep imaging in F090W, F115W, F150W, F200W, F277W, F356W, F410M and F444W was taken on 7th - 8th October, 2022 UT. Shortly after, between 20th - 24th October 2022, deeper imaging in these same bands, plus F335M, was taken as part of GTO 1210 (PI Luetzgendorf), forming Epoch 2. Approximately one month later, GSBD5 was imaged in F182M, F210M, and F444W as part of FRESCO (GO 1895; PI Oesch; Oesch et al. 2023), and F200W+F444W with GO 2198 (Barrufet et al. 2025), both taken on 18th November 2022. We consider both of these programs to be Epoch 3.

Eleven months later, on 14th October 2023, GSBD5 was imaged in F115W, F150W, F200W, F277W, F356W and F444W as part of the pure parallel program PANORAMIC (GO 2514; PIs Williams, Oesch; Williams et al. 2025) as Epoch 4, along with ultra-deep data in medium bands F162M, F182M, F210M, F250M, F300M and F335M taken around the same time from GO 3215 (PI Eisenstein), i.e. the JADES Origins Field (Eisenstein et al. 2025). Finally, a further year later, imaging was taken in F070W, F150W, F200W and F444W as part of GTO 4540 (PI Eisenstein) on 19th October 2024, forming Epoch 5. The programs and observations are detailed in Table 1, and the epoch coverage of each band is tabulated in Table 2.

Band	Flux (nJy)	S/N	Epoch of observation
F070W	$< 0.56$ ($-0.20 \pm0.28$)	0.7	E1
F090W	$< 0.38$ (0.06 $\pm$ 0.19)	0.3	E1, E2, E3
F115W	1.45 $\pm$ 0.34	4.3	E1, E2, E3
F150W	1.39 $\pm$ 0.21	6.5	E1, E2, E3, E4
F162M	3.77 $\pm$ 0.30	12.4	E1, E2
F182M	$< 0.34$ (0.06 $\pm$ 0.17)	0.4	E1, E2, E3
F200W	1.60 $\pm$ 0.37	4.4	E1, E2, E3, E4, E5
F210M	1.86 $\pm$ 0.47	3.9	E1, E2, E3
F250M	$< 0.89$ (0.77 $\pm$ 0.45)	1.7	E1, E2
F277W	8.58 $\pm$ 0.30	29.0	E1, E2, E3, E4
F300M	17.50 $\pm$ 0.38	46.1	E1, E2
F335M	2.77 $\pm$ 0.33	8.5	E1, E2, E3, E4
F356W	61.65 $\pm$ 0.68	90.8	E1, E2, E3, E4
F410M	467.34 $\pm$ 2.01	232.3	E1, E2
F444W	609.54 $\pm$ 2.36	258.2	E1, E2, E3, E4, E5
F560W	360.67 $\pm$ 49.98	7.2	–
F770W	393.54 $\pm$ 42.04	9.4	–
F1000W	1422.17 $\pm$ 80.91	17.6	–
F1280W	1675.82 $\pm$ 152.18	11.0	–
F1500W	2037.08 $\pm$ 182.45	11.2	–
F1800W	1780.58 $\pm$ 480.79	3.7	–
F2100W	$< 1243.52$ (1089.02 $\pm$ 621.76)	1.8	–
F2550W	$< 8651.33$ (2422.85 $\pm$ 4325.66)	0.6	–

We also used Mid-Infrared Instrument (MIRI, Glasse et al. 2015; Wright et al. 2023) imaging data, from SMILES (GTO 1207; PI Rieke; Rieke et al. 2024). These data were taken in December 2022 in F560W, F770W, F1000W, F1280W, F1500W, F1800W, F2100W and F2550W. We used the DR1 mosaic products, the data reduction process for which is detailed in Alberts et al. (2024).

NIRCam imaging for other sources

In addition to GSBD5, we analyse the other nine Y dwarf candidates reported in JADES (Hainline et al. 2026b) in the later section of this paper. For the candidates located in the GOODS-South region, as well as the aforementioned JADES GTO 1180, FRESCO and PANORAMIC datasets, we include imaging from GTO 1286 (PI Luetzgendorf) and DD 6541 (PI Egami), as well as the calibration programs, CAL/NIRCAM 6630 (PI Boyer) and CAL/NIRCAM 8882 (PI Boyer).

For the brown dwarf candidates in GOODS North, we include JADES imaging from GTO 1181 (PI Eisenstein; see Eisenstein et al. 2026) and GTO 1264 (PI Colin Robledo). GOODS North also includes pointings from FRESCO (GO 1895), CONGRESS (GO 3577; PI Egami), GO 2674 (PI Arrabal Haro), TWINKLE (GO 7404; PIs Naidu, Matthee, Chisholm), and the pure-parallel imaging data sets PANORAMIC, SAPPHIRES (GO 6434; PI Egami) and POPPIES (GO 5398; PI Kartaltepe).

Spectroscopic data

GSBD5 was observed with the JWST Near Infrared Spectrograph (NIRSpec; Jakobsen et al. 2022; Böker et al. 2023) using the Micro-Shutter Assembly (MSA; Ferruit et al. 2022) and the low resolution $R \simeq30$–$330$ PRISM setting as part of the OASIS program (GO 5997;

PI Looser). It received 27.7 hours of total exposure time from 2025-01-06 00:37:02.566 UTC (MJD 60681.02572414352) to 2025-01-07 06:57:08.550 UTC (MJD 60682.2896822921). The observations were split between two visits on the two dates, and 36 exposures were obtained in three dithered configurations. The MSA target coordinates were (03:32:20.165, $-27$:50:21.75), i.e., $\sim0.1^{\prime\prime}$ southward and $\sim0.05^{\prime\prime}$ westward from the coordinates observed in October 2022.

To perform spectroscopic data reduction, we use version 1.20.2 of the pipeline (Bushouse et al. 2025) and CRDS_CTX=jwst_1481.pmap version of the JWST CRDS files. First, we acquired the level 1 “uncal” data products from the Mikulski Archive for Space Telescopes (MAST). We then processed these products with the level 1 pipeline (calwebb_detector1) following the methods for MSA slit spectroscopy in Leung et al. (2026), where we used the default configurations for all steps except we turned on the clean_flicker_noise step to remove the effects of 1/f noise from the exposures. For this clean_flicker_noise step, we set fit_method to median, mask_science_region to True, background_method to None and n_sigma to 2. Next, we ran the level 2 pipeline (calwebb_spec2) using the default configurations. We then masked any pixels in the resulting level 2b “cal” files that had at least one flag among a list of selected data quality flags.¹ Finally, we run the level 3 pipeline (calwebb_spec3) using the default configurations to obtain a combined, rectified 2D spectrum, and extract the 1D spectrum via optimal extraction (Horne 1986).

Astrometric measurement

Reference-frame calibrations

Reference calibration is essential for improving JWST/NIRCam astrometry beyond the standard pipeline-calibrated precision and robustly measuring the motion of GSBD5. We first aligned the individual JWST exposures to establish a consistent relative reference frame, and subsequently tied this frame to the International Celestial Reference System (ICRS; Arias et al. 1995). Both steps introduce correlated uncertainties that must be propagated into the motion of GSBD5. All astrometric measurements were performed in a projected common tangent plane, $(\xi, \eta)$, whose coordinate system was initially inherited from the JWST pipeline and subsequently calibrated as described below. Its positive directions point eastwards and northwards, respectively.

Relative calibration

The relative calibration largely follows Liu et al. (2026a), but the present data require two modifications to the workflow. First, the images of GSBD5 were observed by several JWST programs with markedly different footprints (Figure 1), instead of by surveys designed to repeat a common footprint (as was the case for Capotauro; Liu et al. 2026a). A single set of reference sources is therefore not present in every exposure. Secondly, GSBD5 is detected independently in several LW filters, permitting a direct cross-band test of the inferred motion. F444W provides the highest target S/N, the largest number of exposures, and coverage of all five epochs, and is therefore used for the adopted astrometric measurement. The remaining detection LW filters F277W, F300M, F356W and F410M, each with final S/N $> 10$ (see § 4), are retained as independent consistency tests.

Vector (PDF)

Fifty F444W exposures, or frames, were obtained over five epochs, and GSBD5 was significantly detected in every exposure. Reference candidates were detected independently with Source Extraction and Photometry (SEP; Barbary 2018), the Python implementation of Source Extractor (Bertin & Arnouts 1996), with centroids measured by Gaussian windows. We retained detections within 55″ of GSBD5 with 0.2″-radius aperture S/N $\ge8$, 2.1 $\le$ FWHM $\le$ 12 detector pixels, axis ratio $q \ge0.25$, and a clear separation from the target. Anomalous pixels were excluded. At most the 500 highest-S/N candidates were retained in each exposure. These deliberately broad cuts admit stars, compact galaxies, and some extended but core-dominated galaxies, whilst suppressing highly elongated blends and diffuse structures.

The preliminary pipeline-produced World Coordinate System (WCS) was used to initialise the cross-exposure associations. We adopted a reference-source matching radius of 0.06″ and required each source to occur in at least two exposures. This produced 858 initial F444W reference candidates. Seven sources were rejected by visible, large offsets between exposures, which include five Gaia stars (see § 3.2) and two galaxies with noise-affected centroids, leaving 851 reference sources and 11,412 retained per-exposure detections. One remaining Gaia star was removed to further reduce motion-related uncertainty, leaving 850 sources, for which there were 11,371 detections. These partially overlapping detections form a connected network: although they are not all commonly present in every exposure, the overlapping sources connect all five epochs. For the other three GSBD5-detected LW bands, 1408 reference sources contributed $\sim43,000$ detections across 85 exposures, with at least 47 reference detections in every exposure. The distribution of these reference sources is illustrated in Figure 1.

For each exposure, we fitted an affine correction among reference sources,

which accounts for translation, rotation, differential scale, and shear. This allowed us to correct the frames within the simultaneous cross-frame reference-source network.

Reference sources were inverse-variance weighted during the correction fit, with the positional uncertainties reported by SEP. However, the nominal centroid uncertainties can become unrealistically small for bright sources and do not include residual PSF mismatch, morphology, or detector-related errors. We therefore imposed an empirical lower bound on the positional uncertainty of each reference source, $\sigma_{\mathrm{used}} = \max(\sigma_{\mathrm{SEP}}, 3.7\ \mathrm{mas})$. This value was determined from the frame-to-frame scatter of the reference sources measured relative to their mean positions after the first iteration of relative calibration with only $\sigma_{\mathrm{SEP}}$. The scatter has a median of 5.15 mas and a 16% lower bound of 4.77 mas, corresponding to a $4$-$\sigma$ equivalent lower-tail value of $\sim3.7$ mas. This choice makes uncertainties below the floor statistically negligible among the $\sim11,000$ detections. Such a conservative prescription caps exceptionally small values and prevents nominally highest-S/N sources dominating the reference sample. We found our reference-source positional uncertainty in good agreement with the approximately 5-mas bright-source centroid scatter relative to Gaia reported by JADES DR5 (Johnson et al. 2026).

Neighbouring reference sources can share background, PSF, and residual distortion errors, so the $\sim11{,}000$ detections cannot be treated as independent. We found weak evidence for residual spatial coherence on a scale of approximately $9.6^{\prime\prime}$, and adopted a conservative $20^{\prime\prime}$ block size to grid the reference field and group the reference sources in each block. Specifically, to construct these blocks, we drew an independent two-dimensional grid zero point uniformly over $[0,20^{\prime\prime})^{2}$, assigned the 850 sources to the resulting occupied blocks, and resampled the occupied blocks with replacement. The continuously randomised division produced 30–36 occupied blocks over the reference-source-covered region and prevents the uncertainty estimate from depending on one arbitrary placement of the block boundaries. These blocks were used in later bootstrapping as basic units.

Evaluating the affine transformation of each exposure at the position of GSBD5 by bootstrapping realisations, we derived that the relative corrections to a common zero point for the five epochs are $(\Delta\xi_{\mathrm{relative}}, \Delta\eta_{\mathrm{relative}}) = (-10.29, +17.49)$, $(+5.37, -17.76)$, $(+10.91, +0.84)$, $(+1.73, +0.23)$, and $(-7.73, +0.80)$ mas, respectively. Their uncertainties after the correction are $(\delta\xi_{\mathrm{relative}}, \delta\eta_{\mathrm{relative}}) = (0.38, 0.51)$, $(0.44, 0.34)$, $(0.83, 0.64)$, $(0.58, 0.42)$, and $(0.47, 0.45)$ mas. Therefore, the relative calibration substantially improved between-frame target position scattering by roughly an order of magnitude from the native JWST-pipeline alignment. The relative corrections remain correlated between epochs and thus enter the motion measurement. It is propagated to the final measurements in a similar block-level bootstrapping approach, as detailed in § 3.3.

Absolute calibration

The relative solution aligns the exposures and corrects their imperfect sky-to-detector projections, but it remains tied to the initial JWST-pipeline WCS and, effectively, the weighted mean apparent motion of the JWST reference ensemble, which affect our localisation and kinematic analysis of the target in the Milky Way. We therefore calibrated the absolute reference frame using both Gaia stars and distant galaxies. The stellar component supplies the direct connection to the ICRS, whereas the galaxy component constrains the residual uncertainties with a larger-sample anchor.

We first searched Gaia DR3 ([88, 89]) for reference stars within the common F444W footprint. Six stars were found, of which one was rejected because its JWST centroid was compromised by the detector edge. The remaining five stars contributed 53 detections across the 50 F444W exposures. Their Gaia positions, proper motions and parallaxes, together with related uncertainties, were propagated from the Gaia reference epoch to the exact times of the JWST exposures using the parallax factors evaluated for the JWST observer. We measured their JWST positions with the hybrid PSF models (created as in § 3.3.1 but with a 4,500 K blackbody SED representative of a K-type star) and jointly fitted a common 2-D affine transformation.

The small number and Gaia-motion uncertainties due to the faintness of the stars make a Gaia-only tie relatively noisy. We therefore supplemented the stellar tie with distant galaxies, for which the apparent motion should be zero. An extragalactic source was required to satisfy the previous reference-source standards and to appear in at least three epochs. It was also required to match either a reliable 3D-HST (Brammer et al. 2012) source within 0.1″, or a high-quality JADES (Curtis-Lake et al. 2026) NIRSpec spectroscopic redshift within 0.2″, and in either case with $z_{\mathrm{spec}} > 0.1$. Candidates within 0.5″ of a Chandra 7-Ms X-ray counterpart (Luo et al. 2017) were removed to avoid variable Active Galactic Nuclei (AGN). The final sample contains 149 distant galaxies. Each galaxy was assigned its own unknown constant position, and the fit measured whether the ensemble retained a common apparent motion after relative alignment. Note that this is different from the previous relative calibration, because here we deliberately only focused on time-dependent motion to trace the residual motion from Gaia reference stars, not the static imagery alignment. In this construction Gaia supplies the connection to the ICRS, whereas the galaxies constrain time-dependent terms (proper motion and parallax) in that connection and thereby reduce the uncertainty arising from the Gaia motion.

The Gaia and extragalactic zero-point solutions are affected by the affine correction in the relative calibration, as it changes the relative position of reference sources between exposures and thus epochs. We subsequently incorporated this uncertainty in the bootstrapping approach detailed in § 3.3. The derived median absolute calibration is (+0.14 mas yr$^{-1}$, −0.21 mas yr$^{-1}$, +0.13 mas) for the $\xi$- and $\eta$-direction proper motion, and parallax, respectively. Its marginal standard deviations are (0.11, 0.10, 0.44) in the same units.

Motion measurement

Hybrid PSF modelling

A key feature of GSBD5 compared to Capotauro (Liu et al. 2026a) is that it has significantly stronger observed S/Ns. Therefore, the PSF model used to extract its astrometry and photometry must match the level of significance, and thus requires a more careful treatment. Following the approach used by Liu et al. 2026a for Capotauro, which involved generating a simulated PSF model with STPSF v2.2.0 (Perrin et al. 2012, 2014), we used the observed NIRSpec/PRISM spectrum of GSBD5 (see § 5) between 3.5 and 5.3 $\mu$m to weight a 41-wavelength polychromatic PSF and the wavefront-sensing optical-path-difference solution closest to the exposure date. Detector sampling and Fourier propagation were each oversampled by a factor of four, and the model was normalised to unit flux. We searched for GSBD5 around the position reported by Hainline et al. (2026b), allowing the centroid to move by up to five detector pixels (~ 0.31″) from its predicted position. All 50 exposures were measured independently; positions and covariances of images separated by less than two days were subsequently combined.

The simulated PSF model may not perfectly describe the observed PSF. Indeed, we found negative holes in the residual image after PSF subtraction at −8% to −2% level compared to the PSF peak for F444W images in five epochs, implying the simulated PSF having a narrower core than realistic. This is a known tendency of simulated NIRCam PSFs being often slightly sharper (e.g., Morishita et al. 2024; Weaver et al. 2024; Harvey et al. 2025). We therefore constructed a hybrid PSF (Figure 2), which combines both the simulation prediction and a Gaussian-convolution broadening kernel, best matched to the local measurement near GSBD5. We selected 14 point sources within 1′ around GSBD5 by requiring a median S/N > 20 among all exposures, a minimum S/N > 15 in any single exposure and a point-source label by SEP. We approximated their PSF weighting with a 4,500 K blackbody SED, though the choice of SED models has been shown to have a minimal impact on the final PSF (Liu et al. 2026a). We subsequently decided on a broadening convolution kernel by fitting the 14 point sources together with tilted local backgrounds and constructed the hybrid PSF for GSBD5, with the same kernel but the exact previous GSBD5 SED assumptions. The hybrid PSF reduced the $\chi^2$ of the GSBD5 fitting by 14% on average, and 46 out of 50 frames were improved. The previous residual peaks were reduced by 50.0% in absolute flux on average, and became −3% to +3% level across five epochs. These diagnostics motivated adoption of the hybrid model for all later analysis.

Furthermore, we tested a purely empirical F444W PSF which is a median stack of 25 bright stars in the JWST/PRIMER survey over the Ultra Deep Survey (UDS) field. This PSF model has the advantage of ~ 2× larger parent sample size and much brighter reference stars (and thus a more certain profile); however, collecting so many bright stars consequently requires a larger mosaic map, and thus naturally includes errors on local variation, drizzling and other uncalibrated offsets. Therefore, it is intuitively expected to be sub-optimal for our GSBD5 measurement. Indeed, this empirical PSF displayed virtually no improvement compared to the simulated PSF alone for both the residual and the fitted $\chi^2$. We therefore proceeded with the hybrid PSFs. The empirical PSF nevertheless provides a useful test of PSF sensitivity to our astrometric results, which we discuss in Section 3.2.3.

We also constructed hybrid PSFs for F356W, F410M, F560W and F770W, where residuals were visible with a simulated PSF. The hybrid PSFs were derived using point sources identified independently in each band, following the same criteria. The hybrid PSFs improved both the residuals and the target-fitting $\chi^2$. For the NIRCam LW bands, the hybrid PSFs were later used for both astrometry and photometry, while F560W was used for photometry only. For other NIRCam and MIRI bands, the simulated PSF models agree well with the target for photometry, and we therefore retained the STPSF models.

Proper motion and parallax

Using the hybrid PSFs, we measured GSBD5 independently in the 50 aligned F444W exposures. For each exposure, we fitted a $17 \times17$ pixel cutout with a tilted background at the position reported by Hainline et al. (2026b). The same fit simultaneously derived GSBD5 photometry, as detailed in § 4. The fitted detector centroid was transformed to the common $(\xi,\eta)$ tangent plane with the JWST-pipeline WCS. The affine correction of the corresponding exposure was then evaluated and applied to its tangent-plane centroid.

The 50 exposures were grouped into five epochs where separation between exposures was shorter than two days. Within each epoch, the corrected exposure positions were combined, resulting in mean MJDs 59860, 59875, 59902, 60231, and 60603, as displayed in Figure 3. For the astrometric fit, we first evaluated time and the JWST-observer parallax factors at the exact MJD of every exposure and then passed this 50-exposure design through the same projector. Thus, the temporal and parallax information within a multi-day epoch was retained. Specifically, we fitted the positions with a five-parameter model

Vector (PDF)

where $t_0 = 60094$ is the mean time measured in Julian years, $(\mu_{\xi}, \mu_{\eta})$ are the eastward and northward proper-motion components, and a single signed parallax $\varpi$ is shared by the two coordinates. In Figure 3, we decompose the proper motion and parallax components from the joint trajectory; the parallax-only measurements are displayed in Figure 4.

Vector (PDF)

In addition to the pixel-fit covariance, the target position is also affected by local background structure, potential undetected faint neighbours, detector defects, and imperfect source subtraction. We therefore estimated these bias terms with 1,000 coherent injection-recovery tracks in the residual maps of the same 50 exposures. A preliminary five-parameter trajectory, as described by Equations (2) and (3), was firstly fitted and used to subtract the source from the images; then we injected artificial sources coherently in the 50 residual images with the same trajectory model. Each track started from the first exposure in the temporal sequence, uniformly putting a random artificial source in area within $1^{\prime\prime}$ of the GSBD5 position. Injections in later exposures were defined by the trajectory at the exact exposure MJDs and the measured F444W flux of the corresponding image. A track would be re-drawn if it put the artificial source onto an invalid pixel or a nearby bright source in one of the exposures. As a result, we obtained 1,000 successful tracks.

Such injection recovery immediately revealed the different position-measurement uncertainties of individual exposures by comparing the injected positions with the recovered positions. In Figure 3 and Figure 4, we illustrate the positional scattering for each epoch, with exaggerated dispersions for better visibility. It can be seen that Epoch 3 has the largest uncertainty among the five epochs, which can be expected from its smaller number of exposures and their shallowness. In contrast, Epoch 2 provides much more certain positional constraint because of both a larger number of exposures and deeper integrations.

Critically, the injection recovery allowed us to combine uncertainties at different steps: the 20″ reference-source blocks (§ 3.1.1) propagated the relative-calibration uncertainty, and the background perturbation was measured from recovering the injected tracks. Specifically, the reference-source blocks were bootstrapped and every realisation was paired with a random choice of the injection-recovery track. The five parameters were then re-fitted to this combination, and a total of 100,000 bootstrap realisations formed the final five-parameter probability distributions.

As a result, we derived the relative trajectory parameters, the probability distributions of which are slightly non-Gaussian (Figure 5):

Vector (PDF)

with $\mu_{\mathrm{tot},\mathrm{rel}} = 49.64^{+0.26}_{-0.32}\ \mathrm{mas\ yr}^{-1}$. Its residual statistic is $\chi^2 = 6.4$ for five degrees of freedom, or $\chi^2_{\nu} = 1.3$, suggesting an excellent agreement between the model and the data. Indeed, as displayed in Figure 3, the temporal progression of the target position in the first three epochs exhibits very high consistency to the fitted trajectory. If fitting the trajectory with only proper motion, we would instead derive a significantly larger $\chi^2_{\nu} = 18.6$, demonstrating that the trajectory cannot be approximated with a linear motion. Projected separately, the reference-frame term contributes $\sim70\%$ to the standard deviations of proper motion and parallax, whereas the target injection term contributes the remaining $\sim30\%$. This is expected, because for a bright source like GSBD5, the residual noise bias should be trivial.

Finally, we tied the relative result to the ICRS with the Gaia-galaxy calibration. As described in § 3.2, the absolute calibration is also affected by relative calibration, and we thus paired the relative-calibration draw with the absolute-calibration solution. We derived the absolute trajectory:

with $\mu_{\mathrm{tot,abs}} = 49.75^{+0.26}_{-0.30}\ \mathrm{mas\ yr}^{-1}$. The total proper motion is consistent with the $50 \pm20\ \mathrm{mas\ yr}^{-1}$ value reported by Hainline et al. (2024a), but has been now improved to S/N $\approx162$. The $15.7^{+1.4}_{-1.3}\ \mathrm{mas}$ parallax is detected at Gaussian-equivalent S/N $= 11.5$. The slightly smaller uncertainty on the proper motion compared to the relatively calibrated result confirms that the Gaia-galaxy anchor can improve how positional offset is explained by removing temporal motions in reference sources. Transforming the parallax draws separately, we obtained the parallax distance,

We stress that this distance is purely from trigonometric parallax and independent from physical property assumptions of GSBD5.

Binary acceleration and other tests

To test whether GSBD5 is under significant perturbation of a (not yet known) companion object, we also fitted its trajectory with additional binary acceleration, i.e., the third term from Taylor expansion of a Kepler orbit:

where $a_{\xi,i}$ and $a_{\eta,i}$ are the acceleration in two perpendicular directions. However, given our data only sparsely cover five epochs, these acceleration terms are expected to absorb the fitting residual and share degeneracy with the parallax. Indeed, while the previous fitting yielded a nominal total acceleration of 2.4 mas, the new fit has a reduced $\chi_\nu^2 = 0.06 \ll1$, suggesting it is likely a false positive from over-fitting.

An injection-recovery test was then conducted to assess the false-positive rate. In this test, artificial F444W sources (with same fluxes as measured) were injected in the 50 residual images following a pre-known trajectory: the trajectory included only the pre-described fitted proper motion and parallax, with no acceleration term. We then repeated the full astrometric fitting procedure with the Equation–(11) and –(12) model on the injected images to determine whether a spurious acceleration could be recovered by chance. This test, whose corresponding distribution from 1,000 injections is presented in Figure 5, showed that the previously fitted acceleration has only a Gaussian-equivalent significance of $\sim0.94\ \sigma$, and therefore we report a non-detection on binary acceleration. This is consistent with the absence of companions in our later search (§ 6).

We further performed a similar examination for parallax, i.e., forcing a parallax fit to an injected trajectory with only proper motion. No parallax at the same significance was found in 500,000 injections, corresponding to a parallax significance lower limit of at least $> 4.7\ \sigma$, consistent with our measured parallax significance of $11.5\ \sigma$. While raising this floor to $\sim10\sigma$ is unpractical (which would require $\sim10^{29}$ injections), this test confirms the robustness of our detection of parallax for GSBD5. We also tested proper motion with static injections and reached the same significance floor. A larger test with more injections is unnecessary given the visibly secure detection.

Another independent check verified the F444W-derived motion with other bands. We repeated the main measurement with both all 180 LW (i.e., F277W, F300M, F356W, F410M and F444W) exposures and the 130 non-F444W LW exposures, respectively. Each group used the same affine-network construction and the offsets were found generally larger between different bands during the relative calibration. The relative parallaxes are $13.9 \pm1.7$ and $10.5 \pm2.9\ \mathrm{mas}$, respectively. The corresponding proper motions are $(-24.0 \pm0.4, -43.9 \pm0.4)$ and $(-27.1 \pm2.9, -43.2 \pm2.7)\ \mathrm{mas\ yr}^{-1}$. They are all largely consistent within uncertainties, though the non-F444W results were seriously worsened by the absence of final-epoch coverage and the additional uncertainty from wavelength-dependent morphology of reference sources. Therefore, we retain the F444W-only measurements as the final results.

Photometric measurement

We measured the photometry of GSBD5 by performing exposure-level PSF fitting, using the same PSF models and time-dependent trajectory framework as in Section 3. For the five high-S/N detection bands, F277W, F300M, F356W, F410M, and F444W, we allowed the centroid to vary within a five-pixel radius. As described above, the resulting centroids were consistent with the trajectory derived from F444W. For the remaining 19 bands, we performed forced PSF photometry at the predicted position of GSBD5 in each exposure at its exact observation time. Photometric uncertainties were estimated from 1,000 injection-recovery trials on the residual maps.

After subtracting the best-fit PSF, we found no significant residuals in most bands (Figure 6). The only exception is a faint source-like residual in the stack of all 30 F410M exposures, located 0.25″ east and 0.09″ north of GSBD5. A PSF fit to this residual gives $F_\nu= 4.49 \pm1.03$ nJy, corresponding to $\sim1\%$ of the GSBD5 flux in F410M, $m_{\mathrm{AB}} = 29.77$, and S/N = 4.34. No comparable asymmetric residual was produced when the same hybrid PSF was applied to F410M Gaia stars. Moreover, the signal is contributed nearly uniformly by 24 of the 30 F410M exposures, while the remaining six exposures are noise-dominated. These tests suggest that the residual is unlikely to be caused solely by PSF mismatch. Its non-detection in the other bands, especially F444W, argues against a normal Galactic stellar source. Possible interpretations are therefore either an image artefact, or a compact high-redshift emission-line galaxy.

The MIRI photometry is comparatively straightforward, because only the SMILES mosaic products are available and the exposure-level timing information is not retained. In addition, the broader MIRI PSFs make the measurements less sensitive to small positional uncertainties. For the NIRCam data, however, we measured photometry at three time resolutions: individual exposures, individual epochs constructed from exposure-level measurements, and the all-epoch combined photometry. Because brown dwarfs can be photometrically variable (e.g. Radigan et al. 2012; Apai et al. 2013; Metchev et al. 2015; Vos et al. 2020, 2022; Biller et al. 2018, 2024; Sutlieff et al. 2023, 2024; Liu et al. 2024a; McCarthy et al. 2024, 2025; Cushing et al. 2026), combining measurements taken at different epochs may bias the final fluxes. We therefore tested for variability before adopting the stacked NIRCam photometry.

We searched for periodic photometric variability using a Lomb–Scargle (LS, Lomb 1976; Scargle 1982; VanderPlas 2018) analysis of the per-exposure photometry within each epoch. For the high-S/N bands F277W, F300M, F356W, F410M, and F444W, we obtained 17 band–epoch combinations, with temporal baselines ranging from 0.3 to 83 hours and single-exposure times varying by up to an order of magnitude. Thirteen of these provide baselines longer than 2 hours and are therefore suitable for periodicity searches. We required the fitted semi-amplitude to exceed the uncertainty of the corresponding integrated photometry, and required the candidate period to be longer than the break-up limit for a 20 $M_{\mathrm{J}}$ object, $P > P_{\mathrm{break}} = 0.66$ h. Searching periods up to 20 h, we found no physically plausible solution with a significance of $\geq3\sigma$. We further assessed false positives using a control sample of 25 bright, compact high-redshift ($z > 3$) galaxies selected from the reference-source compilation. Owing to cosmological time dilation, non-blazar galaxies (e.g., regular AGN/quasars or compact/bursty star-forming galaxies; Winkler et al. 1992; MacLeod et al. 2016; Izotov et al. 2016; Ricci & Trakhtenbrot 2023; Liu et al. 2024b, 2026b) at these redshifts are not expected to show intrinsic near- to mid-IR variability on the short timescales probed here, and therefore provide a useful empirical control sample for observational systematics. For all tentative LW-band variability signals with nominal significances of 1.0–3.0$\sigma$, the control galaxies produced equal or higher LS powers. We therefore attribute these tentative signals to observational systematics (see e.g. Alderson et al. 2023; Schlawin et al. 2023; Sutlieff et al. 2025; Zhou et al. 2026) and find no significant evidence for periodic photometric variability in GSBD5 within current sensitivity.

We also searched for long-term variability between epochs, and found no $> 5\%$ semi-amplitude between epoch-integrated fluxes. The absence of significant variability justifies combining measurements from different exposures and epochs to obtain the final stacked fluxes reported in Table 2. The control-sample test, however, revealed artificial flux variations at the $\sim1.5\%$–3% level, likely reflecting residual JWST image-to-image zero-point differences in the data used here. We therefore imposed a 5% uncertainty floor on all JWST fluxes in the subsequent analysis. This is consistent with the flux error floor used in Hainline et al. 2026b. Most variable brown dwarfs are observed to vary in the IR at a comparable amplitude (0.2% to 10%, e.g., Metchev et al. 2015; Vos et al. 2022; Biller et al. 2024; Liu et al. 2024a; Cushing et al. 2026). Therefore, we cannot exclude GSBD5 having a weaker variability than that detectable with current flux uncertainties, although the data indeed rule out extreme variability amplitudes, which can sometimes reach as high as almost 40% (e.g., Radigan et al. 2012; Eriksson et al. 2019; Bowler et al. 2020; Zhou et al. 2022; Sutlieff et al. 2023, 2024; Tan et al. 2025).

We then used the NIRCam photometry to calibrate the PRISM spectrum. JWST/NIRSpec PRISM spectra are known to suffer from wavelength-dependent baseline offsets and slit losses, which was particularly important for Copaturo (Liu et al. 2026a). We therefore applied a second-order wavelength-dependent correction by matching synthetic PRISM photometry to the measured fluxes in the strong-detection (S/N $> 5$) bands F150W, F162M, F277W, F300M, F335M, F356W, F410M, and F444W using their filter profiles. The largest correction was required by F277W, at the $\sim8\%$ level, whereas the remaining bands required corrections of only $\leq3\%$. After this calibration, the PRISM spectrum is photometrically consistent with the NIRCam measurements. We note that the Galactic extinction is negligible for GSBD5. At its coordinates, we estimated maximum reddening $E(B-V) = 0.009$ mag (Schlegel et al. 1998; Schlafly & Finkbeiner 2011) with IRSA Galactic DUST Reddening and Extinction service² for an extragalactic object, corresponding to about 1.4% change for F070W flux and 0.1% for F480M. Since GSBD5 is within the Milky Way, the extinction is even less.

Spectrophotometric analysis

Spectral type classification

Hainline et al. 2026a classified GSBD5 as a Y0–Y1 dwarf from its PRISM spectrum, but noted a tension between the SW and LW spectral features. Using the PRISM spectrum recalibrated with our updated photometry, we revisit the spectral classification of GSBD5 in detail.

We first assembled a set of reference spectra spanning late-T to early-Y dwarfs, including the T8 standard 2MASS J04151954−0935066 (Burgasser et al. 2006, data from Burgasser et al. 2004), the T9 standard UGPS J072227.51−054031.2 (Cushing et al. 2011, data from the SpeX Prism Library; Burgasser 2014), the representative T9.5 dwarf WISEA J094306.00+360723.3 (Schneider et al. 2015, data from the Cloud Atlas HST/WFC3 library; Schneider et al. 2015; Manjavacas et al. 2019), the Y0 standard WISEP J173835.52+273258.9 (Cushing et al. 2011, data from the Cloud Atlas HST/WFC3 library; Schneider et al. 2015; Manjavacas et al. 2019), and the tentative Y1 standard WISE J035000.32−565830.2 (Kirkpatrick et al. 2012, data from the Cloud Atlas HST/WFC3 library; Schneider et al. 2015; Manjavacas et al. 2019). To enable a direct comparison with the NIRSpec/PRISM observations, we convolved each reference spectrum to the corresponding instrumental resolution using the msafit³ forward model for a centred point source (de Graaff et al. 2024). The convolved spectra were then normalised to the $J$-band peak flux of GSBD5, as shown in Figure 7.

We first assessed the conventional approach of assigning spectral subtypes based on spectral morphology alone (e.g., Cushing et al. 2011; Kirkpatrick et al. 2012). Specifically, we fitted the $J$-band peak of GSBD5 using the convolved standard spectra. Among the standard templates, the Y1 standard provides the best fit, with a reduced chi-square of $\chi_\nu^2 = 1.2$. To assess the robustness of this result, we independently examined the $J$- and $H$-band peaks in the two NIRSpec visits (Figure 7), both of which show a consistent preference for the Y1 template. Nevertheless, the subtype assignment remains uncertain, as the Y1 template improves $\chi^2$ by only $\sim0.1$ relative to the next-best standard template, with five degrees of freedom (dofs). The $H$-band peak provides even weaker constraints: owing to its lower flux and larger relative uncertainties, the observed profile is consistent with all the reference subtypes considered here.

Notably, the blue wing of the GSBD5 $J$-band peak appears slightly narrower than that of the Y1 standard, a morphological characteristic generally associated with later spectral types. We therefore extended the comparison to the NIRSpec spectra of two of the coldest known brown dwarfs. One is WISE J085510.83−071442.5, hereafter WISE 0855 (data from Luhman et al. 2024); this object is the coldest brown dwarf known to date ($T_{\rm eff} \sim290$ K) and has been tentatively, though somewhat arbitrarily, classified as $\geq$ Y4 (e.g., Kirkpatrick et al. 2019). Another is WISEPA J182831.08+265037.8 (Cushing et al. 2021), hereafter WISE 1828. This object has an exceptionally red colour and once held the record for being the coldest brown dwarf ($T_{\rm eff} \sim300$ K; Kirkpatrick et al. 2012; Beichman et al. 2013). Tentatively classified as $>$ Y2 (Cushing et al. 2021), it has received rich spectroscopic observations, including JWST NIRSpec and MIRI (De Furio et al. 2023; Barrado et al. 2023). However, interpretations of WISE 1828 must be treated with caution because it has been suspected to be an unresolved binary (Leggett et al. 2025; Leggett 2026). Relative to the Y1 standard, WISE 1828 yields a further reduction in $\chi^2 \sim0.9$ for five dofs, while WISE 0855 reduces $\chi^2$ by an additional $\sim0.5$. Thus, in terms of the minimum $\chi^2$, GSBD5 is most closely matched by WISE 0855, the coldest known brown dwarf. However, visual inspection of the $J$- and $H$-band profiles in Figure 7 reveals that the Y0 and later-type spectra are largely indistinguishable within the observational uncertainties. Consequently, spectral morphology alone supports a Y-type classification for GSBD5, primarily based on the blue wing of the $J$-band peak, but does not allow a reliable determination of its subtype.

GSBD5 illustrates a fundamental limitation of spectral classification with JWST NIRSpec/PRISM data. The instrument has its lowest resolving power at short wavelengths, reaching $R_{\rm min} \sim30$ near the $J$-band peak (Jakobsen et al. 2022). As the $J$-band peak becomes progressively fainter and narrower towards later spectral types because of the colder Planck function leaving little intrinsic emission in the Wien tail (Schneider et al. 2016), the combination of limited spectral resolution, increasing relative flux uncertainties, and fewer independent resolution elements makes it increasingly difficult to distinguish between subtypes. In the absence of established spectral standards later than Y1, the classification of objects with unusually narrow $J$-band peaks, such as GSBD5, is poorly defined. We therefore explore alternative empirical proxies for the spectral types of the coldest brown dwarfs, aiming to establish a more practical classification framework for JWST observations and to identify a parameter space in which GSBD5 can be more distinctly characterised.

The absolute magnitude can provide an empirical tracer of spectral type (e.g., Filippazzo et al. 2015). We constructed a library of 539 dwarfs, including 25 M dwarfs from IRTF/SpeX spectral library (Cushing et al. 2005; Rayner et al. 2009; Villaume et al. 2017), 396 L dwarfs and 95 T dwarfs from the SUDA/SPHEREx QR2 (10.26131/IRSA652; Gagné et al. 2026), and 11 T dwarfs and 12 Y dwarfs from Beiler et al. (2024b), excluding WISEA J220905.75+271143.6 because of its noisy SW spectrum. Each object is required to have an unambiguous standard spectral subtype, a measured parallax enabling a reliable flux–luminosity conversion, and a well-calibrated spectrum covering a similar range comparable to that of JWST NIRSpec/PRISM. Additionally, we included the PRISM spectra of WISE 0855 and WISE 1828.

As illustrated in Figure 8, where only a representative subsample of L0–Y4 dwarfs is shown for clarity, the spectral luminosities follow a clear sequence of decreasing luminosities toward later types. Across the full PRISM wavelength range, GSBD5 is significantly fainter than all Y0–Y1 dwarfs in the comparison sample. At $\lesssim4\,\mu\mathrm{m}$, it is only more luminous than WISE 0855, while at $\sim4 - 5\,\mu\mathrm{m}$, the two objects have comparable luminosities. This clearly places GSBD5 at the boundary of, if not well beyond, the luminosity range current defined by Y0–Y1 standards. However, the comparison also reveals an important band dependence in luminosity-based typing. When evaluated at $\lesssim3.5\,\mu\mathrm{m}$, GSBD5 lies between Y1 and the tentative Y4 classification of WISE 0855, but remains substantially more luminous than WISE 0855; when evaluated at $\gtrsim4.2\,\mu\mathrm{m}$, its luminosity, or absolute magnitude, becomes comparable to that of WISE 0855. WISE 1828 also stands out from this sequence with its exceptionally high luminosity at $\gtrsim2.8\,\mu\mathrm{m}$, which has led to the interpretation of an unresolved binary (Leggett et al. 2025; Leggett 2026); conversely, the low luminosity of GSBD5 is in line with it being isolated (see § 6.1 for more discussion).

Vector (PDF)

The overall spectral colour is also known to correlate with the spectral type (e.g., Kirkpatrick et al. 2021; Kiman et al. 2026a). To isolate spectral shape from luminosity, we normalised the 17 JWST T8–Y1 dwarf spectra, as well as GSBD5, WISE 1828 and WISE 0855, by their total NIRSpec/PRISM-integrated flux. The resulting comparison is shown in Figure 9. These cold-dwarf standards reveal a clear spectral sequence from late-T to Y types: the short-wavelength flux decreases systematically toward later spectral types, and this trend reverses at $\gtrsim4\,\mu\mathrm{m}$, although with a more moderate amplitude. This evolutionary sequence is physically expected. As dwarfs cool toward later types, their effective temperatures decrease and the bulk of their emergent flux shifts to longer wavelengths. The increased opacity of H$_2$O, CH$_4$ and NH$_3$ further weakens the SW emission (Faherty & Vos 2026), and the evolving J-peak profile is a by-product of this overall spectral evolution. Overall, GSBD5 lies near the coldest end of the T8–Y1 sequence of the comparison sample across most of the PRISM wavelength range. It shares similar spectral profiles to WISE 1828 at the H peak and $\sim3\,\mu\mathrm{m}$ to $4.5\,\mu\mathrm{m}$, strongly suggesting a subtype no earlier than Y1.

Vector (PDF)

We also note that the spectral evolution becomes more gradual on the LW side toward the coldest objects, causing the later-Y spectra to converge more closely in this wavelength range. This partly reflects the reduced sensitivity of the normalised LW continuum to decreasing effective temperature, but it also indicates the growing importance of additional atmospheric processes at colder temperatures. Indeed, the diversity of spectral features in this region shows that, although effective temperature remains the dominant driver of the subtype sequence, the observed spectra are also shaped by CO$_2$ and CO absorption. The strengths of these features depend on metallicity, gravity, and the degree of vertical mixing in individual objects (Beiler et al. 2024a), producing diverse absorption troughs near $\sim4.3$, $4.6$, and $4.8\,\mu\mathrm{m}$.

Figure 10 shows the F150W–F480M colour evolution with spectral type. A notable feature is that the late-T and Y dwarfs form a smooth, monotonic sequence, without abrupt breaks or reversals. Similar behaviour has been seen in pre-JWST colour–spectral-type diagrams, e.g., Kirkpatrick et al. (2021). We describe this late-T/Y tail phenomenologically with a softened power law,

Vector (PDF)

where $C$ is the colour in AB magnitude, constant $w=0.65$ is the fixed smoothing width, $p$ is the power, $k_1$ is the horizontal offset, and $x=30+q$ for subtype Y$q$ and T$10+q$. To assess how the earlier-type sequence connects to this late-T/Y tail, we further modelled the joint M to T sequence, with a linear profile earlier than late L, and a softened turning curve between late L and early T,

where $x=0-9.5$ for M0–M9.5, $10-19$ for L0–L9, and $20-29$ for T0–T9. The main relation is fitted with all individual sources, with the 1- and 2-$\sigma$ shades marking the dispersion of their colour to the fitted relation for a given spectral type. Sources marked by coloured symbols are not involved in the fitting. We further explored colour combinations of different JWST bands. The full set of colour–type fits, the used data, and the fitted parameters are introduced in Appendix § 9.

This relation only reflects how the colour follows the existing subtype convention, instead of genuine physical correlations. Nevertheless, such a continuous parameterisation is useful because it provides a quantitative proxy for the spectral evolution underlying the classification sequence. Kirkpatrick et al. (2012) discussed which spectral quantity can best distil the progression of the $J$-peak sequence, but at the time the available data favoured classification through direct comparison of $J$-peak morphology. With JWST, the F150W−F480M provides a complementary diagnostic: it traces the same broad late-T/Y spectral evolution as the $J$-peak sequence, while providing a simple and continuous quantity that can be applied uniformly to faint objects. It therefore represents a natural proxy to select late-type

candidates before spectroscopy becomes available, with quantitative predictions into the colder regime beyond the current classification limit. Because of limited reference sources beyond Y1, we extrapolate the uncertainties with those at Y1 in Figure 10.

Applying the fitted relation to GSBD5 gives a colour-predicted spectral type of Y1.5$^{+0.5}_{-0.5}$ (for colour), assuming a 1-$\sigma$ equivalent spectral-type uncertainty. This is consistent with the conclusions from both its spectral luminosity and flux-normalised spectral profile. Interestingly, two sources classified beyond Y1, WISE 0336B and WISE 1828, have F150W−F480M colour nearly identical to that of GSBD5. Together, these three objects may indicate an emerging population beyond the Y1 standards. However, the colour–spectral-type relation remains poorly constrained at the Y0–Y1 end, with fewer than $\sim20$ reference objects, and its behaviour at later subtypes is currently unknown. As shown in Figure 10, the tentative Y4 object WISE 0855 does not follow a simple extrapolation of this sequence. Assigning precise spectral subtypes in this regime therefore remains premature. Nevertheless, the continuous colour relation provides an empirical framework for identifying and comparing objects beyond the current classification standards, which can be progressively refined as more ultracold dwarfs become available for spectroscopic characterisation. We therefore adopt >Y1 as the formal spectral classification of GSBD5, a colour-based proxy of Y1.5, and leave more robust spectral typing to future high-resolution spectroscopy.

Bolometric luminosity

With the parallax measured and therefore distance inferred, the observed SED can be directly converted into an absolute luminosity scale, allowing the bolometric luminosity, $L_{\mathrm{bol}}$, to be estimated with essentially no dependence on physical-property assumptions. The flux shortward of $0.6\,\mu\mathrm{m}$ is negligible (expected $<0.1\%$; Beiler et al. 2023). Over the NIRSpec/PRISM wavelength range, 0.6 to 5.3 $\mu$m, we directly integrated the observed spectrum. At longer wavelengths, from 5.3 to 12 $\mu$m, we fitted the MIRI photometry using the 12 Y-dwarf templates from Beiler et al. (2024b), all of which have MIRI spectroscopic coverage. The measured MIRI fluxes, together with their uncertainties, were directly used in the fitting. We then adopted the median of the fitted templates as the fiducial SED. Beyond 12 $\mu$m, only three templates remain available, i.e., the $>$Y1 candidates, WISE 1828 and WISE 0855, and the Y0 dwarf WISE 1738. Similarly, we fitted these templates to the MIRI photometry, and adopted their median. Beyond 18 $\mu$m, we appended a blackbody tail by fitting it with the 12–18 $\mu$m median, allowing its temperature to vary between 175 K and 500 K. The SED profile is shown in Figure 11. Uncertainties were drawn from Monte Carlo (MC) sampling of the extrapolated spectral templates, PRISM spectral uncertainties, a linear uniform prior of blackbody-tail temperature, and the probability distribution of the parallax.

Vector (PDF)

This procedure gives a bolometric flux $\log_{10}(F_{\mathrm{bol}}/\mathrm{erg\,s^{-1}\,cm^{-2}})=-15.18^{+0.03}_{-0.06}$, a bolometric luminosity $\log(L_{\mathrm{bol}}/L_{\odot})=-7.08^{+0.09}_{-0.09}$ and a blackbody-tail temperature $T_{\mathrm{BB}}=371^{+91}_{-114}$ K. We note that $T_{\mathrm{BB}}$ describes only a component of GSBD5 and it is different from the global description, effective temperature $T_{\mathrm{eff}}$, although they are physically consistent, as in the models discussed below. The PRISM spectrum contributes $\sim22\%$ of the total luminosity, the MIRI-constrained region contributes $\sim58\%$, and the blackbody tail contributes the remaining $\sim20\%$. To assess the sensitivity of $L_{\mathrm{bol}}$ to the poorly constrained 12 − 18 $\mu$m SED shape, we repeated the calculation by omitting the three long-wavelength templates, and appended the blackbody tail directly after the Beiler et al. (2024b) templates at 12 $\mu$m. This approach yielded essentially the same $L_{\mathrm{bol}}$, only $\sim5\%$ higher than the fiducial value. We therefore conclude that the MIRI photometry provides sufficient leverage on the peak of the SED and thus our bolometric luminosity estimate.

Evolutionary modelling

We propagated the full $L_{\mathrm{bol}}$ distribution through two evolutionary models appropriate for cold dwarfs, ATMO2020 (Phillips et al. 2020) and Sonora Flame Skimmer (Mang et al. 2026), to infer the physical properties of GSBD5. We adopted a uniform mass prior, $M\sim\mathrm{U}(0.1,30)\,M_{\mathrm{J}}$, and a log-normal age prior, $\ln(t/\mathrm{Gyr})\sim\mathcal{N}(\ln3.5,[\ln2]^2)$, truncated to $0<t<13.8$ Gyr. The age prior is motivated by the dynamically inferred ages of T dwarfs (Hsu et al. 2021), which provide the closest age-calibrated analogues to cold Y dwarfs. We verified that adopting a uniform age prior, $t\sim\mathrm{U}(0.1,10)$ Gyr, yields essentially unchanged results. Samples falling outside the valid parameter ranges of the evolutionary grids were discarded. For ATMO2020, we evaluated the model luminosity along each $(M,t)$ evolutionary track against the input $L_{\mathrm{bol}}$ distribution by assigning a weight proportional to the luminosity likelihood, i.e., an importance weight. The Sonora Flame Skimmer models comprise 56 evolutionary grids, spanning equilibrium and disequilibrium chemistry, seven metallicities, and four C/O ratios. We only consider the solar-metallicity grids. We performed the same Bayesian analysis within each Sonora grid and assigned equal weight to all grids. The effective temperature $T_{\mathrm{eff}}$, surface gravity $\log g$, and radius $R$ were then obtained from the interpolated model properties. We report the results in Table 3.

Property	Sonora Flame Skimmer	ATMO2020	Unit
$M$	$8.2^{+2.8}_{-2.5}$	$8.5^{+3.2}_{-2.7}$	$M_{\mathrm{J}}$
Age	$3.9^{+2.9}_{-1.8}$	$4.0^{+2.9}_{-1.9}$	Gyr
$R$	$1.04^{+0.03}_{-0.04}$	$1.02^{+0.03}_{-0.04}$	$R_{\mathrm{J}}$
$T_{\mathrm{eff}}$	$301^{+17}_{-16}$	$305^{+18}_{-16}$	K
$\log g$	$4.28^{+0.16}_{-0.18}$	$4.31^{+0.17}_{-0.20}$	cgs

Our importance-weighting approach is mathematically equivalent to exact-likelihood rejection sampling (Dupuy et al. 2023): weighting prior draws by the luminosity likelihood and accepting them with probability proportional to that likelihood target the same posterior distribution. To test it, we conducted the traditional rejection sampling following the same steps as described in Miles et al. (2023). The test yielded essentially the same posterior distributions.

The full posterior distributions of the inferred physical properties are shown in Figure 12. The ATMO2020 and Sonora Flame Skimmer models give broadly consistent results, with median values agreeing to within a few percent for all parameters. The inferred effective temperature is broadly consistent within $1\,\sigma$ with the atmospheric-model estimates of $T_{\mathrm{eff}}=322^{+9}_{-7}$ K using Sonora Elf Owl and $330^{+5}_{-5}$ K using ATMO2020 from Hainline et al. (2026b), although the latter used somewhat different photometry data. GSBD5 therefore remains the coldest brown dwarf identified in the Hainline et al. (2024a) JADES+CEERS sample. These results support the interpretation of GSBD5 as an ultracool $>$Y1-type brown dwarf. Its inferred properties are summarised in Table 3, and we discuss its low mass and broader physical implications in § 6.

Vector (PDF)

Atmospheric modelling

We further explored atmospheric modelling. To explore possible physical scenarios for GSBD5, we collected seven major atmospheric forward model packages, including nine model families and spanning a range of physical treatments. They are:

• $i)$ ATMO2020 (Phillips et al. 2020): a cloud-free grid including both chemical equilibrium (CEQ) and disequilibrium (DEQ) chemistry under different vertical mixing (VM) strengths. We adopted $T_{\mathrm{eff}}=200-600$ K, $\log g=2.5-5.5$, and CEQ/DEQ+weak VM/DEQ+strong VM scenarios.

• $ii)$ ATMO2020++ (Leggett et al. 2021; Meisner et al. 2023): an empirical extension of ATMO2020 to improve cold T/Y dwarf 4–15 $\mu$m SED fitting with DEQ chemistry and a non-adiabatic pressure-temperature relationship. We adopted the same $\log g$ and [M/H] range as for ATMO2020, and $T_{\mathrm{eff}}=200-1200$ K.

• $iii)$ Sonora Bobcat (Marley et al. 2021): a classic cloud-free, rainout-CEQ grid that provides a simple equilibrium baseline. We adopted $\log g=3.25-5.5$, $T_{\mathrm{eff}}=200-600$ K, and [M/H] $=-0.5$, 0, 0.5.

• $iv)$ Sonora Elf Owl (Mukherjee et al. 2024): a cloud-free DEQ model allowing VM, metallicity, and C/O. It was used as the most appropriate Sonora model in Hainline et al. (2026b). We adopted $\log g=3.0-5.5$, $T_{\mathrm{eff}}=275-600$ K, and [M/H] $=-1$ to 1, C/O $=0.5-2.5$, and $\log K_{zz}=2$, 4, 7, 8, 9.

• $v)$ Sonora Flame Skimmer (Mang et al. 2026): the latest cloud-free Sonora model grid, extending to $T_{\mathrm{eff}}=50$ K and $\log g=2.0$, with CEQ/DEQ chemistry and a broad metallicity and C/O coverage. We adopted $\log g=2.0-5.5$, $T_{\mathrm{eff}}=200-600$ K, and [M/H] $=-1$ to 2, C/O $=0.5-2.5$, CEQ/DEQ, and $\log K_{zz}=1-9$.

• $vi)$ Water-cloud model from Lacy & Burrows (2023): this model specifically designed for Y dwarfs allows for water/ice/thick water clouds and EQ/DEQ with carbon and nitrogen molecules, although

with a sparse grid. We adopted $\log g = 3.5 - 5$, $T_{\mathrm{eff}} = 200 - 450$ K, and [M/H] = $-0.5$, 0, 0.5, and thin/thick clouds in combinations with CEQ/DEQ.

vii) Exo-REM k26 (Radcliffe et al. 2026): this family includes three branches with different cloud treatments, i.e., with varying sedimentation efficiency factor $f_{\mathrm{sed}}$, simple microphysics, and cloud-free. The varying $f_{\mathrm{sed}}$ adjusts for condensate particle sedimentation efficiency, although the clouds here refer to silicate and iron clouds, usually at deeper layers of dwarfs and not necessarily dominate Y dwarfs. We adopted $\log g = 3 - 5$, $T_{\mathrm{eff}} = 200 - 500$ K, $f_{\mathrm{sed}} = 1 - 9$, [M/H] = $-0.5$ to 2.0, C/O = $0.1-0.8$ for the varying-$f_{\mathrm{sed}}$ grid, and remove unrelated parameters for the latter two.

Because the Exo-REM k26 family contains three separate branches, we essentially fitted nine model grids in total. All flux uncertainties include the 5% error floor justified in § 4. We used species (Stolker et al. 2020) for grid interpolation and PyMultiNest (Buchner et al. 2014; Buchner 2016) for Bayesian sampling. First, we fitted to the NIRSpec/PRISM spectrum and the MIRI photometry, providing the highest spectral-resolution constraint available for this object, and the results are summarised in Appendix § 10 in Table B1. Second, we used the full NIRCam+MIRI photometric SED, comprising 23 bands in total, with the results tabulated in Table B2. This separation allows us to distinguish failures in reproducing detailed spectral structure from failures in reproducing the broader SED shape, as discussed separately below.

Model

Best branch/config.

$N/k$

$T_{\mathrm{eff}}$ (K)

$\log g$ (cgs)

[M/H] (dex)

C/O

(C/O)/(C/O)$_\odot$

$\log K_{zz}$ (cm$^2$ s$^{-1}$)

$f_{\mathrm{sed}}$

$\log_{10}(R/d)^2$

$d_{1R_J}$ (pc)

$\varpi_{1R_J}$ (mas)

$\chi_\nu^2$

$\ln Z_{\mathrm{family}}$

ATMO2020

weak DEQ

192/3

366.0

3.64

–

–

–

–

–

-22.1699

281.7

3.549

23.142

6910.445 ± 0.111

ATMO2020++

–

192/4

375.2

4.00

0.058

–

–

–

–

-21.6577

156.2

6.40

12.416

7922.118 ± 0.143

Sonora Bobcat

rainout EQ

192/4

533.0

5.25

−0.500*

–

–

–

–

-23.5015

1305.1

0.766

25.430

6710.927 ± 0.096

Sonora Elf Owl†

–

174/6†

405.6

5.50*

−0.500

–

2.077

8.334

–

-22.7358

540.5

1.850

24.427

6400.448 ± 0.144

Sonora Flame Skimmer

DEQ

192/6

380.7

2.03

-0.592

–

0.500*

6.750

–

-21.8458

194.0

5.155

14.467

7736.933 ± 0.161

Exo-REM cloudless

–

192/5

450.0

4.57

−0.49485*

0.800*

–

–

–

-22.8930

647.8

1.544

23.847

6860.989 ± 0.134

Exo-REM simple cloud

–

192/5

500.0*

4.22

−0.49485*

0.796

–

–

–

-23.0896

812.3

1.231

23.734

6871.952 ± 0.134

Exo-REM varying-$f_{\mathrm{sed}}$

–

192/6

500.0*

4.23

−0.49485*

0.800*

–

–

1.000*

-23.0260

754.9

1.325

23.207

6925.268 ± 0.156

Lacy & Burrows (2023)

thin-cloud DEQ

192/4

394.1

3.76

−0.500*

–

–

–

–

-22.2781

319.1

3.134

20.685

7148.074 ± 0.134

Model

Best branch/config.

$N/k$

$T_{\mathrm{eff}}$ (K)

$\log g$ (cgs)

[M/H] (dex)

C/O

(C/O)/(C/O)$_\odot$

$\log K_{zz}$ (cm$^2$ s$^{-1}$)

$f_{\mathrm{sed}}$

$\log_{10}(R/d)^2$

$d_{1R_J}$ (pc)

$\varpi_{1R_J}$ (mas)

$\chi_\nu^2$

$\ln Z_{\mathrm{family}}$

ATMO2020

weak DEQ

23/3

298.2

2.61

–

–

–

–

–

-20.8113

59.0

16.961

21.929

−293.857 ± 0.114

ATMO2020++

–

23/4

315.2

2.50*

-0.721

–

–

–

–

-21.1425

86.3

11.583

10.961

−183.873 ± 0.129

Sonora Bobcat

rainout EQ

23/4

363.7

3.25*

−0.500*

–

–

–

–

-21.5013

130.5

7.664

55.102

−601.316 ± 0.122

Sonora Elf Owl†

–

19/6†

300.0

3.00*

-0.158

–

1.297

8.943

–

-20.9399

68.4

14.627

25.496

−222.703 ± 0.148

Sonora Flame Skimmer

DEQ

23/6

310.9

2.00*

-0.254

–

0.909

9.000*

–

-20.8711

63.2

15.832

5.984

−136.350 ± 0.142

Exo-REM cloudless

–

23/5

264.2

3.00*

-0.414

0.800*

–

–

–

-20.6226

47.4

21.076

22.043

−280.590 ± 0.136

Exo-REM simple cloud

–

23/5

264.6

3.00*

-0.417

0.800*

–

–

–

-20.6269

47.7

20.972

21.923

−280.037 ± 0.137

Exo-REM varying-$f_{\mathrm{sed}}$

–

23/6

301.8

3.00*

-0.284

0.800*

–

–

1.000*

-20.8632

62.6

15.978

5.467

−136.301 ± 0.157

Lacy & Burrows (2023)

thick-cloud DEQ

23/4

276.1

3.75

0.000

–

–

–

–

-20.6604

49.6

20.179

23.041

−302.584 ± 0.135

and cloudless Exo-REM, which yield the largest $\chi_{\nu}^{2}$ values and the lowest Bayesian evidence values, ln $Z$ (see Table B1). All three are cloud-free models.

The best fits are obtained with ATMO2020++ and Sonora Flame Skimmer. However, they imply distance-to-radius ratios, $d/R$, that correspond to equivalent distances of $d_{\mathrm{1RJ}} \sim150$–200 pc for a typical Y-dwarf radius of $R = 1\,R_{\mathrm{J}}$, in > $20\sigma$ tension with the parallax-derived value. We find a similar discrepancy compared to fits to the PRISM spectrum alone by Hainline et al. (2026a), who obtained $\chi_{\nu}^{2} = 19.63$ and $d = 139$ pc with ATMO2020++, and $\chi_{\nu}^{2} = 34.85$ and $d = 188$ pc with Sonora Elf Owl. The recurrence of this discrepancy suggests that the problem is not specific to a single model family, but instead reflects a broader mismatch between current atmospheric models and the observed spectrum of GSBD5.

Another, perhaps more fundamental, feature is that the spectrophotometric fits systematically favour high effective temperatures (see a discussion of this feature in Carter et al. 2023b). The fitted temperatures have a mean of $T_{\mathrm{eff}} = 434$ K and a median of 406 K, substantially higher than the evolutionary-model estimate of $\sim307$ K. A similar temperature bias has been reported for WISE 0855 (Luhman et al. 2024). Note that for a fixed observed bolometric flux and radius, the inferred distance scales as $d \propto T_{\mathrm{eff}}^{2}$; increasing $T_{\mathrm{eff}}$ from 300 K to 400 K alone therefore increases the inferred distance by a factor of $\sim1.8$. These cases indicate that forward modelling of high-quality spectra for such ultracool dwarfs can bias the inferred temperatures and consequently the distances to higher values, when a $\sim1\,R_{\mathrm{J}}$ radius is assumed. We discuss possible origins of this bias in § 6.3.

SED fitting with PRISM + MIRI

None of the joint NIRSpec/PRISM and MIRI spectrophotometric fits provides a statistically acceptable solution. Across all model families, the reduced chi-square values are large, with a median $\chi_{\nu}^{2} = 23.2$, a mean $\chi_{\nu}^{2} = 21.3$, a standard deviation of 4.4 and a minimum of 12.3. The poorest fits are obtained with Sonora Bobcat, Sonora Elf Owl,

SED fitting with NIRCam + MIRI

We fitted all 23 NIRCam+MIRI photometric bands listed in Table 2 using the same nine model families with native flux measurements capped at the 5% uncertainty floor. The resulting parameters are summarised in Table B2. Overall, the fits remain statistically poor: the reduced chi-square values have a median of $\chi_\nu^2 = 21.9$ and a mean of $\chi_\nu^2 = 21.3$. The large standard deviation of $\sigma_{\chi^2} = 14.0$ is driven primarily by the Sonora Bobcat fit.

Nevertheless, several models reproduce the photometric SED substantially better than they reproduce the joint PRISM+MIRI spectrophotometry. The best solutions are obtained with two models, Exo-REM varying-$f_{\rm sed}$ and the Sonora Flame Skimmer DEQ, both with $T_{\rm eff} \sim300$ K and $d_{\rm1RJ} \approx63$ pc. These inferred temperatures and flux scalings are broadly consistent with the parallax distance and with the effective temperature inferred from evolutionary models. These results are also similar to those reported in Hainline et al. (2026b). Fitting with selected NIRCam and MIRI bands, they estimated $d = 70$ pc and $T_{\rm eff} = 322$ K with Sonora Elf Owl, and $d = 90$ pc and $T_{\rm eff} = 330$ K with ATMO2020. Thus, while the models still fail to reproduce the full photometric SED at high precision, the photometric fits recover broadly reasonable global SED profile and flux scaling.

We then repeated the fitting after removing all NIRCam medium bands. The Exo-REM varying-$f_{\rm sed}$ and Sonora Flame Skimmer models again provide the best fits, with little change in fit quality.

This indicates that the models perform similarly at the NIRCam medium+wide-band resolution, corresponding to an effective spectral resolution of $R \sim10$–$15$, and at the wide-band-only resolution of $R \sim4$. In other words, reducing the data to even broader photometric bands does not substantially alleviate the remaining mismatch.

The relative performance of different model families provides useful clues about the atmospheric physics favoured by GSBD5. Within the Exo-REM family, the varying-$f_{\mathrm{sed}}$ model fits much better than either the simple-cloud or cloud-free branches. The latter two also require less satisfactory flux scalings, with $d_{1\mathrm{RJ}} \sim47$–$48$ pc. This suggests that adding clouds alone is not sufficient; rather, additional freedom in the cloud sedimentation efficiency, or a sedimentation-like modification of the vertical atmospheric structure, is necessary. The same preference for DEQ and/or vertical mixing is also confirmed by the significant improvement within the Sonora family from Bobcat to Elf Owl and finally Flame Skimmer.

Prior-guided SED fitting

The parallax measurement and evolutionary modelling provide independent constraints on the physical properties of GSBD5. A natural next step is therefore to use these constraints as priors in the SED fitting. However, since most atmospheric models only have very few (degenerate) parameters than those obtained through the evolutionary modelling, constraining all known properties via priors would compound parameter compensation in the fitting. We therefore only input the full Sonora Flame Skimmer $T_{\mathrm{eff}}$ and $\log g$ posteriors in Figure 12 as priors, and allowed other properties to vary. Critically, the free-fitted flux scaling factor, $q = (R/d)^2$, can separate the spectral morphology from the overall flux scaling, allowing us to assess whether the fitted SED remains consistent with the parallax-based distance and expected radius.

whether the fitted SED remains consistent with the parallax-based distance and expected radius. We adopted the joint two-dimensional prior distribution obtained by equally combining the Sonora Flame Skimmer and ATMO2020 evolutionary-model posteriors (see § 5.3 and Figure 12). We fitted all nine model branches to the 23-band NIRCam+MIRI photometry using the same fitting setup as above, except for the inclusion of the evolutionary-model priors. The resulting SEDs, together with the NIRCam and MIRI photometry and the PRISM spectrum for comparison, are shown in Figure 13. The fitted parameters are listed in Table B3.

Vector (PDF)

Model	Best branch/config.	$N/k$	$T_{\mathrm{eff}}$ (K)	$\log g$ (cgs)	[M/H] (dex)	C/O	(C/O)/(C/O)$_\odot$	$\log K_{zz}$ (cm$^2$ s$^{-1}$)	$f_{\mathrm{sed}}$	$\log_{10}(R/d)^2$	$d_{1R_J}$ (pc)	$\varpi_{1R_J}$ (mas)	$\chi^2_\nu$	$\ln Z_{\mathrm{family}}$
ATMO2020	strong DEQ	23/3	284.8	3.417	–	–	–	–	–	-20.8091	58.8	17.003	35.784	−444.621 ± 0.128
ATMO2020++	–	23/4	350.0	4.048	0.285$^{*}$	–	–	–	–	-21.1546	87.5	11.424	13.672	−211.580 ± 0.150
Sonora Bobcat	rainout EQ	23/4	335.3	3.768	0.004	–	–	–	–	-21.3120	104.9	9.530	67.088	−720.048 ± 0.135
Sonora Elf Owl$^{\dagger}$	–	19/6$^{\dagger}$	277.9$^{*}$	3.427	0.143	–	0.934	8.983$^{*}$	–	-20.7186	53.0	18.871	42.062	−336.335 ± 0.189
Sonora Flame Skimmer	DEQ	23/6	288.6	3.420	0.223	–	0.817	8.302	–	-20.7533	55.2	18.131	31.599	−362.292 ± 0.189
Exo-REM cloudless	–	23/5	281.0	3.415	−0.454$^{*}$	0.789$^{*}$	–	–	–	-20.8645	62.7	15.953	34.280	−399.259 ± 0.179
Exo-REM simple cloud	–	23/5	281.0	3.415	−0.457$^{*}$	0.797$^{*}$	–	–	–	-20.8682	63.0	15.884	34.318	−400.167 ± 0.180
Exo-REM varying-$f_{\mathrm{sed}}$	–	23/6	281.1	3.414	−0.403$^{*}$	0.800$^{*}$	–	–	1.001$^{*}$	-20.8453	61.3	16.309	32.744	−379.052 ± 0.210
Lacy & Burrows (2023)	thick-cloud DEQ	23/4	276.2	3.750	0.000	–	–	–	–	-20.6645	49.8	20.083	23.046	−304.556 ± 0.160

*†Same as in Table B1 and Table B2.

The fitted global parameters in the prior-guided fits change in a physically meaningful way. Solar or subs olar metallicities stay largely unchanged, and the new flux scalings remain similar, corresponding to a median distance of $\sim61$ pc for $R = 1 R_{\mathrm{J}}$. The effective temperatures drop to $\sim281$ K (i.e., lower than the $\sim307$ K prior), and the gravity estimate is improved from the previous grid lower boundaries to a median of $\log g \sim3.42$.

The large $\chi^2_\nu$ values are dominated by a small number of bands. As shown in Figure 13, F115W, F277W and F300M consistently show $\gtrsim10\,\sigma$ discrepancies across nearly all model branches, whereas the residuals in the other bands vary more strongly from model to model. For a typical number of degrees of freedom of $\sim20$, these three bands alone increase the reduced $\chi^2_\nu$ by about 15. Both F277W and F300M are assigned the 5% flux-error floor because of their high native S/N. The large residuals therefore mainly reflect highly significant relative discrepancies with respect to the adopted uncertainties. This highlights the importance of high-S/N PSF photometry across a complete set of bands. With lower-S/N measurements, such as the $\sim3\times$ higher F277W noise adopted for GSBD5 by Hainline et al. (2026b), or with missing bands of any of the three, the statistical significance of this mismatch would be substantially reduced.

Controlled SED-fitting comparison experiments

Because no single atmospheric model family provides a statistically acceptable fit to the SED of GSBD5, we performed a set of controlled $\chi^2$ comparison experiments to identify which additional model ingredients most effectively reduce the mismatch. These experiments are designed to compare the relative statistical performance of models that differ in specific assumptions and to identify missing ingredients in current models, a strategy that has been widely used in cold-dwarf studies when no fully adequate model is available (e.g., Morley et al. 2014; Leggett et al. 2017; Miles et al. 2020; Lacy & Burrows 2023; Beiler et al. 2024b).

The current best-performing free-grid model, Exo-REM with variable $f_{\mathrm{sed}}$, does not include water clouds. However, water clouds are expected to become important in atmospheres with $T_{\mathrm{eff}} \sim300$ K (e.g., Morley et al. 2014; Lacy & Burrows 2023). We therefore tested the effect of water clouds using different branches of the Lacy & Burrows (2023) model family. We fitted six combinations of clear, thin-cloud, and thick-cloud atmospheres, each with CEQ and DEQ chemistry, to the same NIRCam+MIRI photometry over the same parameter ranges. Within this model family, the best fit is obtained with the thick-cloud DEQ model, followed by the thin-cloud DEQ and thick-cloud CEQ models. All cloudy branches perform better than the corresponding clear branches, and the DEQ models generally outperform their equilibrium counterparts. The same qualitative preference is obtained when fitting the PRISM+MIRI data. Thus, within the Lacy & Burrows (2023) framework, the SED of GSBD5 favours both water clouds and DEQ chemistry, plausibly associated with vertical mixing. However, the current SED fitting cannot uniquely identify if this preference comes from the cloud composition or cloud structure or both.

We next tested whether horizontal inhomogeneity can improve the fit. Physically, this experiment represents an atmosphere composed of multiple columns with different local thermal or cloud properties (e.g., Chen et al. 2025b; Oliveros-Gomez et al. 2026). Using Exo-REM with variable $f_{\mathrm{sed}}$, we fitted the NIRCam+MIRI photometry with $M = 1$–5 atmospheric components. The components were required to share the same $\log g$, [M/H], and C/O ratio, but were allowed to have different $T_{\mathrm{eff}}$ and $f_{\mathrm{sed}}$. Increasing the number of components produces only a modest improvement in Bayesian evidence, at the $\sim7\%$ level, with the best case occurring at $M = 2$. Even this best multi-component fit still has $\chi_\nu^2 > 5$. We therefore find that horizontal patchiness alone cannot explain the SED mismatch.

We then explored a complementary possibility that the mismatch reflects an incomplete description of the vertical atmospheric structure. Different wavelengths probe different pressure levels (e.g. Buenzli et al. 2012; Biller et al. 2013, 2024; Ge et al. [94]; Vos et al. 2023; McCarthy et al. 2024, 2025; Chen et al. 2025b; Oliveros-Gomez et al. 2026), so an imperfect pressure–temperature profile, opacity structure, or vertical mixing prescription could appear observationally as a wavelength-dependent redistribution of emergent flux. To test this phenomenologically, we introduced a smooth wavelength-dependent weighting function, $w(\lambda)$, that reshapes the relative SED brightness. We parameterised $w(\lambda)$ with polynomials of order $N = 0$–5. The case $N = 0$ corresponds to no wavelength-dependent correction beyond the usual flux scaling, while $N = 1$ allows a linear fit across wavelength. This experiment improves the Bayesian evidence by $\lesssim15\%$ across the tested polynomial orders (Figure 14). Thus, a smooth wavelength-dependent correction moderately improves the fit, but higher-order corrections provide diminishing returns.

Vector (PDF)

Finally, we combined horizontal multi-component structure with wavelength-dependent redistribution. We modelled the observed flux as

where $q = (R/d)^2$ is the overall flux-scaling factor, $w_m(\lambda) \ge0$, and $\sum_{m=1}^{M} w_m(\lambda) = 1$ at each wavelength. Here, $F_{\lambda,m}$ represents the spectrum of the $m$-th atmospheric component, while $w_m(\lambda)$ describes its wavelength-dependent contribution to the disk-integrated emission. Combinations with excessively large $(M,N)$ were omitted when the number of free parameters exceeded the information content of the available photometric data. The resulting Bayesian evidences are shown in Figure 14.

The comparison across $(M,N)$ combinations points to a relatively simple phenomenological solution. A large gain occurs at $M = 2$ and $N = 1$, corresponding to two atmospheric components whose relative contributions vary linearly with wavelength. This model gives $\chi^2 = 28.0$ for 13 degrees of freedom, or $\chi^2_\nu= 2.2$, a substantial improvement over all single-component fits. Beyond this point, adding more components or higher-order wavelength dependence produces only marginal gains. For example, increasing the polynomial order changes $\ln Z$ by $< 3\%$, while models with $M = 4$ or 5 components have lower Bayesian evidence despite their additional flexibility. This plateau indicates that the improvement is not simply a numerical consequence of adding free parameters; rather, most of the recoverable mismatch is physically captured by a two-component model with a smooth wavelength-dependent redistribution. We show this preferred $M = 2$, $N = 1$ fit in Figure 15, list the fitted parameters in Table 4, and discuss its physical implications in § 6.3.

Vector (PDF)

Property	Value	Unit
$T_{\mathrm{cool}}$	$244^{+5}_{-8}$	K
$T_{\mathrm{warm}}$	$425^{+14}_{-10}$	K
$T_{\mathrm{total}}$	$284^{+10}_{-8}$	K
$f_{\mathrm{sed,cool}}$	$5.0^{+2.7}_{-2.7}$	–
$f_{\mathrm{sed,warm}}$	$1.03^{+0.04}_{-0.02}$	–
$\log g$	$3.21^{+0.17}_{-0.13}$	cgs
[M/H]	$-0.41^{+0.05}_{-0.05}$	–
C/O	$0.77^{+0.02}_{-0.03}$	–
$d_{1\mathrm{RJ}}$	$52.4^{+4.4}_{-3.6}$	pc
$\varpi_{1\mathrm{RJ}}$	$19.1^{+1.4}_{-1.5}$	mas

Discussion

An isolated planetary-mass object in the thin disk

We searched for possible companions to GSBD5 over a wide range of projected separations. In the immediate vicinity of GSBD5, no plausible point-source companion is detected in the JWST residual images. Given the analysed field of view, this rules out obvious companions over projected separations of $\lesssim7.5 \times10^{3}$ AU, depending on direction (see Figure 1). Because the available imaging spans 0.7–25.5 $\mu$m and reaches very faint limits, a bound stellar companion that is undetected in all bands is strongly disfavoured. Indeed, we cannot rule out the possibility of a similarly cold, low-mass brown dwarf companion at a very small separation, given that the F444W PSF full-width-half-maximum (FWHM) corresponds to a projected distance of $\sim10$ AU, larger than a typical expected Y-dwarf binary separation of a few AU (e.g., (Fontanive et al. 2023)). However, the binary rate of such extremely cold, late-type dwarfs is statistically low (e.g., $\lesssim10\%$; (Fontanive et al. 2018)).

We then searched for a possible wide stellar host. We queried the Gaia DR3 catalogue ([88, 89]) within $1^\circ$ of GSBD5, corresponding to $2.3 \times10^{5}$ AU at the distance of GSBD5. Among the 13 sources with parallaxes compatible with GSBD5, the closest proper-motion match differs by $\sim40.0$ mas yr$^{-1}$. Conversely, the closest overall proper-motion match has a parallax of $\varpi= 5.137 \pm0.013$ mas, in $8.2\sigma$ tension with GSBD5. We therefore find no evidence of a stellar host.

We further searched CatWISE2020 ((Marocco et al. 2021)) and the All-Sky NOIRLab Source Catalogue DR2 ((Nidever et al. 2021)) for fainter M/L/T/Y-dwarf companions. To reduce contamination from compact galaxies and other extragalactic sources, we cross-matched the candidates with the VIDEO DR5 ((Jarvis et al. 2013)) and DES DR2 ([1]) catalogues. We required the proper motion to be consistent with that of GSBD5 within $3\sigma$ and restricted the search to a $5'$ radius, corresponding to $\sim2 \times10^{4}$ AU at the distance of GSBD5. The only remaining candidate is CatWISE J033228.63–275029.5. However, its proper motion is constrained only at the $\sim1\sigma$ level, and its DES colours, compared with the colour templates of dal Ponte et al. (2023), are consistent with an M4 star. Assuming a conservative absolute magnitude of $M_J = 9.5$ mag for an M4 dwarf, its photometry places it at $> 300$ pc, ruling out physical association with GSBD5. We therefore find no plausible dwarf companion to GSBD5. This is consistent with the non-detection of binary acceleration.

With the inferred mass of $\sim8\,M_J$ (§ 5.3), GSBD5 lies at the extreme low-mass end of the brown-dwarf population, far below the hydrogen-burning limit of $\sim75\,M_J$ ((Dieterich et al. 2014; Chabrier et al. 2023)). As a branching threshold for the planetary-mass regime, we adopt $13\,M_J$ ((Spiegel et al. 2011)), below which 95% and 93% of the GSBD5 mass posterior lies, from the evolutionary fits with Sonora Flamme Skimmer and ATMO2020, respectively. Together with the lack of known companions, this places GSBD5 in the regime of isolated planetary-mass objects, or sometimes termed as sub-brown dwarfs or free-floating planets ((Miret-Roig 2023)).

The coordinates place GSBD5 at Galactic coordinates of approximately $(l,b) = (223.62^\circ, -54.47^\circ)$. At the parallax distance of $d \simeq63.7$ pc, this corresponds to a vertical offset of $d \sin b \simeq-51.3$ pc below the Galactic plane, well within the scale height of the Galactic thin disk. This spatial location is consistent with the current kinematic constraints: the measured proper motion implies a tangential velocity of $v_{\mathrm{tan}} = 15.0^{+1.4}_{-1.3}$ km s$^{-1}$, within, and toward the low-velocity end of, the thin-disk ultracool-dwarf tangential velocity distribution (e.g., (Faherty et al. 2009)). However, the current NIRSpec/PRISM spectrum does not provide a useful radial-velocity constraint, so the full three-dimensional space motion remains unknown. Future higher-resolution spectroscopy would be needed to constrain the full kinematics of GSBD5.

Expanding the astrometric frontier for the coldest brown dwarfs with JWST

Until now, no brown dwarf in a JWST extragalactic deep field has had a securely measured parallax, with previous attempts remaining unsuccessful (e.g., (Liu et al. 2026a)). GSBD5 therefore provides the first such case, enabling a route to characterise distant brown dwarfs in extragalactic fields with parallaxes. In the pre-JWST era, all-sky WISE searches provided a flux-limited Y-dwarf sample only out to $\sim10$ pc ((Kirkpatrick et al. 2014, 2016; Wright et al. 2019; Kirkpatrick et al. 2021)). One of the most distant parallax-measured Y-dwarf candidates, the photometric Y0 candidate CWISEP J135937.65–435226.9, has $\varpi= 41.1 \pm4.3$ mas, corresponding to $d \sim24.3$ pc ((Marocco et al. 2021)). The previously most distant spectroscopically confirmed Y dwarf, WD 0806–661B, lies at a parallax-based distance of $\sim19.2$ pc, which was only discovered by targeted, deep Spitzer observation ((Luhman et al. 2011)), and its parallax is measured from its brighter white-dwarf companion ((Subasavage et al. 2009; Luhman et al. 2012; Martin et al. 2018; Fontanive et al. 2026)). The parallax measurement of GSBD5 therefore extends geometric distance measurements of Y dwarfs by a factor of $\gtrsim3$. Hainline et al. (2024a, 2026b) reported 10 Y-dwarf candidates from two of the deepest JWST extragalactic fields (Figure 16), the JADES GOODS-South and GOODS-North fields, including GSBD5. These candidates were selected through SED fitting with inferred $T_{\mathrm{eff}} < 500$ K, but no astrometric motion was reported for any candidate other than GSBD5. Following the same approach used for GSBD5, we analysed the per-exposure photometry and astrometry of the remaining candidates using the reductions described in Section 2. Compared to Hainline et al. (2026b), we now have 10 new NIRCam bands (F070W, F162M, F182M, F187N, F210M, F250M, F430M, F460M, F480M) and slightly different flux values of existing bands, because of both a more complete collection of imaging data than JADES DR5 and the PSF photometry approach. The PSF-based photometric results are tabulated in Appendix D and their astrometry is summarised in Table 5.

ID	$\mu_{\xi}$ (mas yr$^{-1}$)	$\mu_{\eta}$ (mas yr$^{-1}$)	$|\mu|$ (mas yr$^{-1}$)	S/N$|\mu|$	$\varpi$ (mas)	S/N$_{\varpi}$	$d_{\varpi}$ (pc)	$d_{\mathrm{SED}}$ (pc)	$v_{\mathrm{tan},\varpi}$ (km s$^{-1}$)	$v_{\mathrm{tan,SED}}$ (km s$^{-1}$)
JADES-GN-BD-5	$-21.1^{+1.9}_{-1.2}$	$-5.2^{+1.8}_{-2.3}$	$22.0^{+0.7}_{-2.0}$	20.2	$0.0^{+4.9}_{-6.9}$	0.0	$>203$	$870^{+210}_{-210}$	$>21.1^{+0.7}_{-1.9}$	$88^{+23}_{-21}$
JADES-GN-BD-8	$-21.8^{+2.6}_{-2.6}$	$7.4^{+2.9}_{-3.7}$	$23.1^{+2.9}_{-3.0}$	7.7	$1.4^{+1.9}_{-1.7}$	0.8	$>309$	$610^{+80}_{-100}$	$>33.8^{+4.3}_{-4.4}$	$66^{+14}_{-13}$
JADES-GN-BD-9	$-9.9^{+1.9}_{-3.9}$	$1.9^{+3.3}_{-4.4}$	$10.5^{+5.2}_{-2.0}$	3.4	$33^{+11}_{-20}$	2.0	$>23$	$1190^{+330}_{-370}$	$>1.1^{+0.6}_{-0.2}$	$60^{+31}_{-22}$
JADES-GN-BD-11	$-19^{+10}_{-12}$	$18^{+11}_{-15}$	$28^{+13}_{-8}$	2.1	–	–	–	$2090^{+950}_{-1010}$	–	$258^{+203}_{-131}$
JADES-GN-BD-16	$88^{+125}_{-138}$	$213^{+192}_{-95}$	$233^{+192}_{-91}$	1.7	–	–	–	$960^{+450}_{-310}$	–	$1082^{+982}_{-486}$
JADES-GS-BD-4	$10.4^{+4.8}_{-7.0}$	$-1.0^{+5.1}_{-6.6}$	$11^{+6}_{-4}$	1.8	–	–	–	$1380^{+380}_{-420}$	–	$66^{+52}_{-36}$
JADES-GS-BD-8	$12.8^{+3.0}_{-2.4}$	$-11.3^{+1.8}_{-1.6}$	$17.4^{+2.2}_{-2.3}$	8.5	$-0.9^{+1.9}_{-2.0}$	$-0.4$	$>963$	$920^{+150}_{-190}$	$>79^{+10}_{-11}$	$74^{+18}_{-16}$
JADES-GS-BD-21	$9^{+10}_{-13}$	$15^{+9}_{-10}$	$23^{+5}_{-14}$	1.8	$-65^{+62}_{-48}$	$-1.04$	–	$900^{+310}_{-210}$	–	$92^{+48}_{-48}$
PANO-GS-BD-18	–	–	–	–	–	–	–	$440^{+90}_{-140}$	–	–

We robustly detect four additional proper motions at S/N $> 3$, for JADES-GN-BD-5, JADES-GN-BD-8, JADES-GN-BD-9, and JADES-GS-BD-8. These motions confirm that these sources are Galactic objects. We also obtain tentative proper-motion detections at $1.5 < \mathrm{S/N} < 3$ for JADES-GN-BD-11, JADES-GN-BD-16, JADES-GS-BD-21, and JADES-GS-BD-4. The only Y-dwarf candidate without a proper-motion constraint is PANO-GS-BD-18, which has only one available observing epoch.

No additional candidate shows a robust parallax detection with S/N $> 3$. The most suggestive case is JADES-GN-BD-9, with a tentative S/N $\sim2$ parallax of $\varpi\sim33.3$ mas, corresponding to $d \sim30$ pc. However, this parallax is driven almost entirely by the March 2023 epoch, which was observed in F410M rather than the F444W band used in the other epochs. Omitting one of the three exposures from this F410M epoch changes the inferred parallax by $\sim\pm1\sigma$. We therefore do not use this parallax as a direct distance constraint. Instead, using the 68% intervals of all meaningful parallax constraints, we report trigonometric distance limits in Table 5 for the four candidates whose parallaxes can be effectively constrained. These limits are broadly consistent with the SED-based distances inferred from the median of the three models used by Hainline et al. (2026b), although most of them remain relatively loose. The other five candidates either have too few temporal samplings to fit a proper motion + parallax model, or are too noisy for the parallax to be constrained.

Using the SED-fitted distances, the newly measured proper motions correspond to tangential velocities below 100 km s$^{-1}$ for most sources, consistent with the tangential-velocity distribution of thin-disk cold dwarfs (e.g., Faherty et al. 2009). JADES-GN-BD-11 and JADES-GN-BD-16 have formally larger tangential velocities of $\sim260$ km s$^{-1}$ and $\sim1080$ km s$^{-1}$, respectively. However, these values have $\sim80\%$–$100\%$ uncertainties due to the uncertain SED-fitted distances. The current data therefore do not warrant a detailed kinematic interpretation of these two objects.

We also fitted the PSF photometry of the nine candidates using late-T/Y JWST/NIRSpec templates (Beiler et al. 2024b) over the entire NIRCam wavelength range. All candidates are formally best matched by Y-type templates (Table D1). In the colour–magnitude diagrams shown in Figure 17, they lie on the branch occupied by late-T and early-Y dwarfs. However, most of these objects have only upper limits in F150W and F200W, affecting 6/9 and 7/9 sources, respectively. We therefore cannot locate their exact positions on the colour–magnitude diagrams or determine whether they are typical early-Y dwarfs or later-type objects similar to GSBD5. We further explored the colour proxy for spectral types using F356W and F444W, as shown in Appendix A1. These candidates span an approximate range from cT9.4 to cY2, but with large uncertainties owing to the coarser spectral-type sensitivity of F356W-based colours. More spectroscopy is needed to robustly classify this population.

Vector (PDF)

Two objects in this sample, JADES-GN-BD-5 and JADES-GN-BD-9, have been accepted for JWST/NIRSpec observations in Cycle 5 GO 10264 (PI Finkelstein). Based on their F356W and F444W colours, they correspond to approximate continuous types of cT9.4 and cY1.1, respectively. If these objects are late-T or early-Y dwarfs, their spectra, especially at shorter wavelengths, are unlikely to be informative as their NIRCam fluxes are $\sim20$ times fainter than those of GSBD5. We also expect them to be more massive, as their NIRCam fluxes indicate several times higher luminosity given the SED-fitted distances and the fact that the luminosity evolution of Y dwarfs depends more on mass than age (Burrows et al. 2001).

Regardless of whether these two dwarfs can be spectroscopically classified, JWST extragalactic surveys are now revealing a growing population of ultracool brown dwarfs, e.g., Capotauro in the CEERS field (Gandolfi et al. 2026a; Liu et al. 2026b) and two Y dwarfs recently discovered in the Bullet Cluster field (Bradac et al. 2026). High Galactic-latitude extragalactic fields therefore provide a (perhaps unexpectedly) powerful discovery space for Galactic ultracool dwarfs: their exceptional depths, broad filter coverage, and frequent revisits together create favourable conditions for both selection and characterisation.

The main selection challenge in these fields is distinguishing ultracool dwarfs from compact high-redshift galaxies. Recently, cold dwarfs have been mostly confused with LRDs (e.g., (Tu et al. 2025b)). Indeed, GSBD5 itself satisfies all the colour-based LRD selection criteria in Kokorev et al. (2024), Barro et al. (2024) and Labbe et al. (2025). More robust selection techniques (e.g., Kocevski et al. (2025)) requires a photometric redshift check, which can only be obtained with at least two long-wavelength detections and informative non-detections at shorter wavelengths. In other words, an extragalactic survey must guarantee two detections of an object at $\gtrsim3\ \mu\mathrm{m}$ to separate the coldest brown dwarfs effectively from galaxies and form a useful Galactic survey volume.

This criterion allows us to estimate the capability of existing JWST surveys to select these objects. In the following estimate, we conservatively require a brown dwarf to be detected in F277W; for a typical Y dwarf in a JWST deep extragalactic survey, this consequently suggests detections in F356W and F444W, i.e., three detections. NIRCam F277W is comparable to the blue side of the WISE W1 band, and AllWISE, the major pre-JWST survey yielding a large sample of brown dwarfs, had a point-source depth of $m_{\mathrm{AB}} = 19.60$ mag in W1 (Wright et al. 2010; Cutri et al. 2013). In JADES DR5 (Johnson et al. 2026; Robertson et al. 2026), the GOODS-South field covers 210 arcmin$^{2}$ to an aperture-corrected 5-$\sigma$ F277W point-source depth of 30.07 mag, while GOODS-North covers 115 arcmin$^{2}$ to a depth of 29.61 mag. Relative to the AllWISE W1 depth, these limits increase the distance reach by factors of $\sim100$–124. Thus, despite the small solid angle, the JADES search corresponds to a WISE-equivalent survey volume of $\simeq1.43 \times10^{5}$ deg$^{2}$, or $\sim3.5$ times the AllWISE all-sky volume for comparable objects. Since WISE-era all-sky searches identified only a couple of isolated planetary-mass brown dwarfs later than Y1, and a randomly placed source has a $\sim20\%$ probability of overlapping with a background galaxy in the JADES DR5 segmentation map, this simple volume-scaling argument predicts $\sim5.6 \pm5.0$ such objects in the JADES volume, consistent with the discovery of GSBD5.

GSBD5. Considering WISE 0855 is at $\sim2$ pc, such objects will have a characteristic distance of a few hundred pc in extragalactic fields. Deep JWST extragalactic surveys therefore provide an efficient route to expanding the rare sample of the coldest brown dwarfs, as shown in Figure 16. For example, PRIMER adds $\sim380$ arcmin$^{2}$ across COSMOS and UDS with NIRCam, corresponding to roughly two thirds of the WISE all-sky volume at its F277W depth (Liu et al. 2026b; McLeod et al. 2026). COSMOS-Web (Casey et al. 2023) and COSMOS-3D (GO 5893, PI Kakichi) may contribute a further $\sim0.6$–1.5 WISE-equivalent volumes. Assuming a median F277W point-source limit of 28.2 mag, the 0.63 deg$^{2}$ compilation of existing JWST extragalactic fields from McLeod et al. (2026) corresponds to $\sim5$–10 AllWISE all-sky Y-dwarf volumes. These fields could plausibly yield $\sim10$–20 isolated extremely cold planetary-mass brown dwarfs. Such a sample would transform the currently sparse late-Y regime into a statistically useful population. This estimate is conservative, because survey volume scales steeply with limiting magnitude, $V \propto10^{0.6\Delta m}$, and the $\Delta m \sim3.5$ depth range among these surveys implies a $\gtrsim100$-fold variation in searchable volume per pointing.

The implication is not simply a larger search volume. Since WISE Y dwarfs can be detected out to $\sim10$ pc, deep JWST imaging can in principle detect comparable objects out to $\sim1,000$–2,000 pc. At high Galactic latitudes, this distance range corresponds to vertical heights beyond the thin-disk scale height. Although it suggests a reduced effective survey volume towards the Galactic outskirts, it allows a new opportunity to identify ultracool brown dwarfs in Galactic environments that were inaccessible in the WISE era, including the transition between the thin and thick disks, and even the halo.

The main requirement for placing these distant objects on the same physical footing as nearby WISE sources is robust astrometry, especially parallax. The JWST/NIRCam F444W PSF (Nardiello et al. 2022) is $\sim12$–14 times sharper than that of the Spitzer/IRAC 4.5 $\mu\mathrm{m}$ channel (Fazio et al. 2004), partially compensating for the smaller parallax amplitudes at larger distances. In principle, with a follow-up strategy analogous to the Spitzer parallax programs for WISE discoveries, JWST can extend parallax measurements to distances of order $\sim200$–300 pc for favourable targets. The six-month roll-angle flips commonly used in deep extragalactic fields naturally provide useful parallax baselines, and the serendipitous revisits of GSBD5 demonstrate the feasibility of this approach. In the most frequently revisited fields, parallaxes may already be measurable at larger distances than GSBD5, as will be explored in future work (Liu et al., in preparation). Future progress will therefore require three complementary efforts: larger JWST NIRCam and MIRI deep-field imaging areas to build the sample, dedicated revisits of known candidates to constrain their parallaxes, and deep NIRSpec spectroscopic follow-up to reveal their physical conditions.

The atmosphere of ultra-cool giant-planet analogues

GSBD5 in the colour–magnitude diagram

Figure 17 places GSBD5 in colour–magnitude space together with a comprehensive compilation of M–Y dwarfs and directly imaged exoplanets. The first panel shows a traditional ground-based near-IR colour–magnitude diagram based on 2MASS $J$ and $K_{\mathrm{s}}$ photometry, supplemented by synthetic photometry where spectra are available. In this diagram, late-T and Y dwarfs form a sharply turning sequence, with the Y dwarfs occupying a sparsely populated red/faint branch. GSBD5 lies around the faintest and reddest end of this branch, where data are very sparse.

The emergence of such ultracool planetary-mass objects is now becoming clearer in the JWST era. This is illustrated by the three JWST colour–magnitude diagrams shown in Figure 17: F150W−F444W and F200W−F444W, commonly used in extragalactic blind surveys (and exoplanet imaging surveys, e.g. (Carter et al. 2023a, 2024; Biller et al. 2025; Ferrer-Chavez et al. 2026)), and F150W−F480M, often used in JWST ultracool brown-dwarf studies (e.g., (Calissendorff et al. 2023; Albert et al. 2025)). In these colour–magnitude spaces, cold substellar objects form clear, monotonic sequences. GSBD5 lies near the coolest end of the Y-dwarf population, more extreme than some other extremely cold brown dwarfs, such as WISE 1828, WISE 0336B, WD 0806B and WISE 1534, and close to WISE 0855. It is also similar to a few of the coldest directly imaged exoplanets or

candidates: ε Ind A b, with $T_{\mathrm{eff}} \simeq275$ K and $M \simeq7.6 M_{\mathrm{J}}$ (Matthews et al. 2024, 2026, not illustrated here due to lack of data); 14 Her c, with $T_{\mathrm{eff}} \simeq275$–300 K and $M \simeq7.2$–$7.6 M_{\mathrm{J}}$ ([12]); and TWA 7 b, with $T_{\mathrm{eff}} \simeq320$ K and $M \simeq0.3$–$0.4 M_{\mathrm{J}}$ (Lagrange et al. 2025; Crotts et al. 2025). These exoplanets orbit bright K-type host stars at angular separations of only $\simeq4.1^{\prime\prime}$, $\simeq1.1^{\prime\prime}$, and $\simeq1.5^{\prime\prime}$, respectively; and the absence of a coronagraphic spectroscopic mode of JWST means that light from the host star can strongly contaminate the planetary spectrum. Indeed, even for ε Ind A b with the largest star–planet angular separation among the three, recent deep JWST NIRSpec/IFU observations were not able to produce a separate spectrum of the planet (Kravtchenko et al. 2026).

Isolated, self-luminous ultracool objects like GSBD5 avoid this contrast limitation and therefore provide a cleaner route to studying atmospheric conditions analogous to those of mature cold giant planets (e.g., Dupuy & Kraus 2013; Beichman et al. 2014; Faherty et al. 2014; Skemer et al. 2016; Leggett et al. 2019). Although GSBD5 need not have formed through a planet-like pathway, its physical properties place it close to the atmospheric parameter space of mature giant planets. With its high-quality NIRSpec spectrum and MIRI photometry, GSBD5 provides a precious isolated spectroscopic planetary-mass benchmark for studying $\sim300$ K hydrogen-rich super-Jovian atmospheres, where only no more than a few similar sources are known (e.g., WISE 0855; $T_{\rm eff} = 285$ K, $M = 6.5 \pm3.5\ M_{\rm J}$; Wilson et al. 2025).

Physical interpretations of GSBD5 atmosphere

Previous studies of late-T and Y dwarfs have shown that atmospheres in this temperature regime differ substantially from those of warmer L and T dwarfs (see the review by Faherty & Vos 2026). As $T_{\rm eff}$ decreases, molecular opacity from H$_2$O, CH$_4$, and NH$_3$ becomes increasingly important, while the spectra remain highly sensitive to metallicity, gravity, vertical mixing, and cloud formation. DEQ chemistry can dredge CO and CO$_2$ from deeper, hotter layers into the observable atmosphere, strengthening their absorption features (see a discussion on the case of WISE 0855 in Kühnle et al. 2025). At $T_{\rm eff} \lesssim300$ K, liquid water clouds, and potentially water-ice clouds, are also expected to emerge after silicate/iron clouds and subsequently salt/sulfide clouds have retreated to deeper atmospheric layers (e.g., Morley et al. 2014; Lacy & Burrows 2023). WISE 0855 provides an important precedent in this regime, showing evidence for water depletion at mid-IR wavelengths, strong vertical mixing traced by CO (Luhman et al. 2024), and patchy water clouds (Miles et al. 2026).

Our fits provide several indications that vertical mixing is important in the atmosphere of GSBD5. Nearly all model comparisons prefer DEQ chemistry, whether the fits use PRISM+MIRI spectrophotometry, NIRCam+MIRI photometry, or evolutionary-model priors. The evidence for modified cloud sedimentation is more model-dependent, but within the Exo-REM family the fits favour variable-$f_{\rm sed}$ models over simpler cloud prescriptions. These trends suggest that both vertical chemical transport and cloud structure are important for reproducing the SED of GSBD5.

The spectral morphology supports this interpretation. In CEQ, carbon in the upper atmospheres of the coldest Y dwarfs should reside primarily in CH$_4$ (Lodders & Fegley 2002; Zahnle & Marley 2014). Strong CO, and excess CO$_2$, absorption at observable pressures therefore requires DEQ chemistry, most naturally through vertical mixing (e.g. Visscher & Moses 2011; Miles et al. 2020; Luhman et al. 2024; Beiler et al. 2024b; Miles et al. 2026). Consistent with this expectation, GSBD5 shows strong CO and CO$_2$ absorption features around $\sim4.2$–$5\ \mu$m (Figure 9). At the same time, the main CH$_4$ absorption regions, near $\sim1.7\ \mu$m and $3.2$–$3.8\ \mu$m, appear relatively weakened. However, it is unclear if H$_2$O, or indeed water clouds, plays a key role in the atmosphere of GSBD5. We see relatively stronger emission at $1.4$, $1.85$ and $2.5$–$3.0\ \mu$m in GSBD5 (Figure 9), corresponding to water absorption peaks. These absorption troughs that are prominent in other cold Y dwarfs are partially filled in, and in some cases appear as local peaks, even compared to the coldest brown dwarf WISE 0855. However, the low resolution of the PRISM spectrum, as well as its high noise, makes it difficult to identify spectral features exclusively from water, as these bands are also primarily affected by CH$_4$ and NH$_3$. No absorption feature suggesting water ice clouds at $\sim2.8\ \mu$m (Morley et al. 2014) is identified in the PRISM spectrum.

Therefore, we cannot conclude if there is reduced H$_2$O absorption. However, we note that water deficiency would help the SED-fitting of GSBD5, as it naturally produces the extra emission in F277W and F300M, mitigating the model–data discrepancies there found by our prior-guided fits (§ 5.4.3).

In the meantime, the SED-fitting failures can also be mitigated by the two-component fitting approach, even without the involvement of water clouds (§ 5.4.4). In this scenario, the warmer component, with $T_{\rm eff} \sim425$ K, dominates the total flux at $\lesssim3.5\ \mu$m and helps reproduce the relatively weak H$_2$O and CH$_4$ absorption features. This component also favours very low $f_{\rm sed}$, i.e., inefficient sedimentation. However, a single warm, strongly mixed component would produce too much CO absorption at $\sim4$–$5\ \mu$m. The fit therefore shifts the long-wavelength contribution toward a cooler component with more moderate cloud sedimentation. Here, having two different-temperature atmospheric columns is not enough. The (M, N) test reveals the need to tweak the wavelength-dependent structure of these columns, compared to typical brown-dwarf DEQ models.

Therefore, interestingly, here we witness a similar ambiguity previously seen in the interpretations of WISE 0855’s atmosphere (see a summary in Kühnle et al. 2025). On the one hand, water depletion, and even water ice clouds, were favoured by theoretical models for WISE 0855 (e.g., Morley et al. 2014; Lacy & Burrows 2023; Miles et al. 2026). On the other hand, the same data could be similarly interpreted with cloudless models (e.g., Luhman et al. 2024; Kothari et al. 2026). Further investigating the two scenarios of GSBD5 requires bespoke modelling, which is beyond the scope of this paper, or indeed future deep spectroscopy (e.g., Miles et al. 2026).

Causes of GSBD5 misinterpretations

The features described in the section above provide a plausible explanation for the SED-fitting failures discussed in § 5.4.1. When the high-S/N PRISM spectrum dominates the likelihood, the atmospheric fits prefer hotter models, typically $T_{\rm eff} \sim350$–$400$ K, to align with the significant spectral features at NIRCam LW. Because the correspondingly boosted luminosity then needs to match the observed flux, these hotter solutions naturally lead to distances that are much larger than the parallax value. By contrast, NIRCam photometric fitting averages over the detailed spectral structure and places more weight on the broad SED shape, leading to lower temperatures and more reasonable distance estimates. In the latter case, the remaining failures are more associated with limitations of the model grids, such as insufficient temperature coverage of the cold-Y regime (e.g., the LOWZ models; Meisner et al. 2021; Hainline et al. 2026b) or the absence of DEQ chemistry to reproduce the carbon-bearing features at low temperature (e.g., Sonora Bobcat).

Several other fitted parameters can be understood as compensating for missing opacity. In the NIRCam+MIRI photometric fits, many models prefer the lowest surface gravities allowed by their grids, which reduces the photospheric pressure and weakens water and CH$_4$ absorption. They also prefer very high C/O ratios, which can reduce the available oxygen for H$_2$O formation; and high log $K_{zz}$, which enhances vertical mixing, increases the abundance of CO and CO$_2$ and similarly reduces the available oxygen (Visscher & Moses 2011; Moses et al. 2013; Zahnle & Marley 2014; Hobbs et al. 2019; Miles et al. 2020). The extreme combinations in the fits are unlikely to represent the true physical state of GSBD5. For example, extremely low gravity is inconsistent with the evolutionary-model constraints, and the combination of very low metallicity and extremely high C/O is difficult to reconcile with a low-tangential-velocity, likely thin-diskobject. A mildly subsolar metallicity may still be physical, as most models do not require the grid-edge solutions.

At present, however, we do not have a single self-consistent model grid that simultaneously includes DEQ chemistry, sufficiently low temperatures, a broad range of metallicities and C/O ratios, and multiple cloud species including water condensates. Therefore, interpreting physical properties of the coldest objects, especially those distant ones found in extragalactic fields without parallax measurements, needs particular caution. Nevertheless, our analysis provides a clear target for future model development: any successful model for GSBD5 must simultaneously reproduce its phenomenological features.

Conclusions

We have presented a JWST astrometric and spectrophotometric analysis of JADES-GS-BD-5, an extremely cold Y dwarf in the JADES/GOODS-South field. Our main conclusions are as follows:

(i) We obtained the first robust parallax measurement of a brown dwarf discovered in a JWST extragalactic deep field. Using NIRCam images across five epochs spanning a 2.03-yr temporal baseline, we measure parallax $\varpi= 15.7^{+1.4}_{-1.3}$ mas, detected at S/N = 11.5. This gives a model-independent trigonometric distance of $d = 63^{+6}_{-5}$ pc. Its distance is roughly three times larger than that of the previously most distant Y dwarfs with reliable geometric distances. This demonstrates that JWST deep-field imaging can provide highly precise astrometry for extremely faint Galactic substellar objects far beyond the local WISE-discovered sample.

(ii) JADES-GS-BD-5 is likely isolated and belongs to the Galactic thin disk. The absolute proper motion is measured with very high precision, $\mu_{\mathrm{tot}} = 49.75^{+0.26}_{-0.30}$ mas yr$^{-1}$, corresponding to a tangential velocity of $v_{\mathrm{tan}} = 15.0^{+1.4}_{-1.3}$ km s$^{-1}$. Its Galactic position and low tangential velocity are consistent with a thin-disk object. We find no evidence for a close companion in residual images, no significant binary acceleration, no convincing photometric variability and no plausible widely-separated stellar or substellar host.

(iii) JADES-GS-BD-5 is one of the coldest, faintest and latest-type Y dwarfs known to date. Although its short-wavelength $J$-peak morphology alone is broadly compatible with an early-Y classification, its absolute spectral luminosity, flux-normalized PRISM spectrum, and JWST colour–type relation all place it beyond the current Y0–Y1 standards. We therefore adopt a formal classification of > Y1, with a continuous colour-based proxy of cY1.5.

(iv) The parallax-calibrated luminosity places JADES-GS-BD-5 in the planetary mass regime. Integrating the NIRSpec/PRISM spectrum, MIRI-constrained mid-IR SED, and long-wavelength blackbody tail yields $\log(L_{\mathrm{bol}}/L_{\odot}) = -7.08^{+0.09}_{-0.09}$. Evolutionary modelling subsequently gives $M \sim8\,M_{\mathrm{J}}$, $T_{\mathrm{eff}} \sim300$ K, $R \sim1.0\,R_{\mathrm{J}}$ and $\log g \sim4.3$. More than 90% of the mass posterior lies below $13\,M_{\mathrm{J}}$, placing the object at the extreme low-mass end of the brown-dwarf population. It is therefore a rare analogue of mature gas giant planets and one of the only isolated planetary-mass brown dwarf benchmarks to study their atmospheres.

(v) JWST extragalactic surveys provide a new route to the coldest Galactic substellar population. Applying similar astrometric analyses to other nine JADES Y-dwarf candidates confirms eight additional proper motions, although no other robust parallax has yet been detected. The combination of deep 1–5 $\mu$m imaging, repeated visits, and sharp astrometry makes JWST deep fields uniquely powerful for discovering and characterising late-Y dwarfs and isolated planetary-mass objects beyond the local Solar neighbourhood. An ultracool dwarf sample several times larger than that in the WISE era is expected in the existing JWST extragalactic fields. Future dedicated observations will be crucial for turning these serendipitous discoveries into a statistically useful, parallax-calibrated sample.

(vi) We investigated different physical scenarios to explain the challenge for current atmospheric models to reproduce its high-quality near-to-mid-IR SED. Free fits to the NIRSpec/PRISM spectrum and MIRI photometry produce formally unacceptable solutions and tend to overestimate the effective temperature and distance. Photometric fits with NIRCam and MIRI recover more plausible temperatures and flux scalings, but still show significant residuals. Prior-guided and variable-controlled fitting experiments, as well as spectral features, suggest that DEQ chemistry and vertical mixing are needed to explain the observed spectrum of this ultra-cool atmosphere. However, it is unclear if water deficiency happens in GSBD5, as similar SED behaviour can be reproduced by modifying non-water-cloud atmospheric models.

Acknowledgements

DJM and JSD acknowledge the support of the Royal Society through the award of a Royal Society University Research Professorship to JSD. XC acknowledges funding from the China Scholarship Council (CSC) under grant CSC no. 202208170018. BJS and BAB acknowledge funding by the UK Science and Technology Facilities Council (STFC) grant nos. ST/V000594/1 and UKRI1196. HL and ACC acknowledge support from a UKRI Frontier Research Grantee Grant (PI Carnall; grant reference EP/Y037065/1).

This work made use of the University of Edinburgh’s Institute for Astronomy research computing cluster, cuillin, which is partially funded by the UK STFC. We would like to thank Erick Tittley for his work managing and maintaining cuillin. This work is based in part on observations made with the NASA/ESA/CSA James Webb Space Telescope. The data were obtained from the Mikulski Archive for Space Telescopes at the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-03127 for JWST. These observations are associated with program nos. GTO 1180, GTO 1181, GTO 1189, GTO 1207, GTO 1210, GTO 1286, GTO 1275, GO 1895, GO 2198, GO 2514, GO 2674, GO 3215, GO 3486, GO 3577, GO 3670, GO 4539, GTO 4540, GO 4668, GO 5118, GO 5997, GO 6084, GO 6434, DD 6541, CAL/NIRCAM 6630, GO 7404 and CAL/NIRCAM 8882. The authors acknowledge the GO 1895, GO 2514, GO 3215, GO 3577, GO 6434, DD 6541, CAL/NIRCAM 6630, GO 7404, and CAL/NIRCAM 8882 teams for developing their observing programs with a zero-exclusive-access period. This work has made use of data from the European Space Agency (ESA) mission Gaia (https://www.cosmos.esa.int/gaia), processed by the Gaia Data Processing and Analysis Consortium (DPAC, https://www.cosmos.esa.int/web/gaia/dpac/consortium) (Gaia Collaboration et al. 2016, 2023). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the Gaia Multilateral Agreement. This publication makes use of data products from the Spectro-Photometer for the History of the Universe, Epoch of Reionization and Ices Explorer (SPHEREx), which is a joint project of the Jet Propulsion Laboratory and the California Institute of Technology, and is funded by the National Aeronautics and Space Administration. This research has made use of the Spanish Virtual Observatory (SVO) Filter Profile Service “Carlos Rodrigo” https://svo2.cab.inta-csic.es/theory/fps/, funded by MCIN/AEI/10.13039/501100011033/ through grant PID2023-146210NB-I00 (Rodrigo et al. 2012; Rodrigo & Solano 2020). This research made use of SAOImageDS9, a tool for data visualization supported by the Chandra X-ray Science Center (CXC) and the High Energy Astrophysics Science Archive Center (HEASARC) with support from the JWST Mission office at the Space Telescope Science Institute for 3D visualization (Joye & Mandel 2003). Some of the data products presented herein were retrieved from the Dawn JWST Archive (DJA). DJA is an initiative of the Cosmic Dawn Center (DAWN), which is funded by the Danish National Research Foundation under grant DNRF140. This research has made use of the Astrophysics Data System, funded by NASA under Cooperative Agreement 80NSSC21M00561. This work makes use of the Python programming language⁴, in particular packages including Astropy [8, 9, 10], Matplotlib (Hunter 2007), Numpy (Harris et al. 2020), Photutils (Bradley et al. 2026), Scipy (Virtanen et al. 2020), species (Stolker et al. 2020), PyMultiNest (Buchner et al. 2014; Buchner 2016), sep (Barbary 2018), and STPSF (Perrin et al. 2012, 2014).

Data availability

The data underlying this article will be shared on reasonable request to the corresponding author.

Colour-driven spectral typing

We constructed the empirical colour–spectral-type relations from an enlarged parent library of 899 spectroscopically classified objects. The parent library comprises 25 M0–M9.5 spectra from the IRTF/SpeX Spectral Library (Rayner et al. 2009); 845 L0–T7 SPHEREx Quick Release 2 spectrophotometric spectra from the SPHEREx Ultracool Dwarf spectrophotometric Atlas (SUDA; Tu et al. 2026); 23 T6–Y1 JWST spectra from Beiler et al. (2024b); three T5–T8 JWST/NIRSpec spectra from the RUBIES brown-dwarf sample (Morrissey et al. 2026); two independently classified T5 and T8 JWST/NIRSpec standards from Pritchard et al. (2026), for which we adopted the original DAWN v4.4 reductions; and direct F150W and F480M photometry of 20 Y dwarfs and the T9 standard UGPS J072227.51−054031.2 from Albert et al. (2025) (NIRCam photometry is used in priority to the synthesized photometry).

Synthetic AB photometry was calculated by integrating each spectrum through the JWST/NIRCam system-throughput curves obtained from the SVO Filter Profile Service (Rodrigo et al. 2012, 2024). The sample was selected independently for each colour: an object was retained only when it had positive measured fluxes, raw $S/N > 2$ in both constituent bands, and sufficient spectral coverage in both filters. Consequently, the fitted sample varies between panels, from 216 to 425 objects, while the union of all objects entering at least one relation contains 720 unique sources.

The two objects historically assigned the intermediate Y0.5 subtype in the Beiler et al. (2024b) sample were excluded from all fits and shown only as external diagnostics. WISE J033605.05−014350.4B (Albert et al. 2025), WISEPA J182831.08+265037.8 (Cushing et al. 2021; De Furio et al. 2023), WISE J085510.83−071442.5 (Luhman et al. 2024), and GSBD5 (Hainline et al. 2026b) were likewise displayed only for comparison and did not constrain the fitted relations. The WISE 1828+2650 colours were synthesized from its public JWST/NIRSpec PRISM spectrum obtained under program GTO 1189 (PI: T. Roellig), whereas the GSBD5 colours were derived from the recalibrated PRISM spectrum presented in this work. We illustrate the relations in Figure A1, tabulate the fitted parameters in Table A1, and summarise the criteria for different spectral types in Table A2.

Vector (PDF)

Colour	$N$	$a$	$s_1$	$s_2$
F150W − F480M	245	−1.839 [−1.871, −1.811]	0.108 [0.104, 0.111]	−0.129 [−0.160, −0.057]
F150W − F460M	313	−1.790 [−1.812, −1.749]	0.105 [0.103, 0.109]	−0.097 [−0.148, −0.068]
F150W − F444W	309	−1.647 [−1.672, −1.614]	0.123 [0.119, 0.126]	−0.043 [−0.100, −0.025]
F150W − F410M	372	−1.512 [−1.535, −1.478]	0.135 [0.130, 0.137]	−0.056 [−0.094, −0.031]
F200W − F480M	291	−1.629 [−1.648, −1.593]	0.079 [0.074, 0.080]	−0.136 [−0.092, 0.036]
F200W − F460M	388	−1.572 [−1.584, −1.524]	0.075 [0.070, 0.077]	−0.005 [−0.033, 0.026]
F200W − F444W	358	−1.434 [−1.451, −1.402]	0.093 [0.089, 0.094]	0.039 [−0.001, 0.075]
F200W − F410M	425	−1.234 [−1.289, −1.235]	0.094 [0.093, 0.103]	0.111 [0.042, 0.114]
F277W − F480M	216	−1.005 [−1.034, −0.973]	0.053 [0.051, 0.057]	0.069 [0.020, 0.080]
F277W − F460M	289	−0.973 [−0.990, −0.931]	0.054 [0.049, 0.056]	0.035 [−0.084, 0.033]
F277W − F444W	277	−0.796 [−0.826, −0.781]	0.064 [0.064, 0.069]	0.122 [0.092, 0.131]
F277W − F410M	336	−0.653 [−0.676, −0.634]	0.075 [0.073, 0.077]	0.127 [0.102, 0.142]
F356W − F480M	221	−0.580 [−0.601, −0.548]	−0.008 [−0.011, −0.006]	0.023 [0.009, 0.175]
F356W − F460M	294	−0.516 [−0.550, −0.500]	−0.012 [−0.013, −0.008]	0.070 [−0.023, 0.107]
F356W − F444W	295	−0.375 [−0.396, −0.350]	0.004 [0.001, 0.006]	0.050 [0.046, 0.130]
F356W − F410M	367	−0.216 [−0.238, −0.211]	0.012 [0.011, 0.014]	0.056 [0.051, 0.085]
Colour	$A$	$p$	$k_1$	$k_2$
F150W − F480M	0.022 [0.009, 0.099]	2.559 [2.016, 2.939]	18.613 [17.639, 19.060]	21.326 [21.000, 22.871]
F150W − F460M	0.089 [0.010, 0.260]	2.098 [1.642, 2.971]	18.282 [17.302, 18.860]	23.242 [21.000, 24.279]
F150W − F444W	0.063 [0.007, 0.253]	2.234 [1.647, 3.000]	18.390 [18.063, 19.266]	23.589 [21.407, 24.747]
F150W − F410M	0.013 [0.006, 0.127]	2.761 [1.887, 3.000]	18.564 [18.414, 19.453]	22.295 [21.247, 24.239]
F200W − F480M	0.140 [0.041, 1.163]	1.878 [1.020, 2.362]	20.708 [17.896, 20.629]	22.708 [22.034, 26.395]
F200W − F460M	0.258 [0.026, 1.111]	1.661 [1.020, 2.515]	17.115 [16.764, 18.184]	24.220 [21.994, 26.123]
F200W − F444W	0.738 [0.163, 1.176]	1.205 [1.020, 1.893]	19.636 [18.620, 21.000]	26.046 [24.516, 26.888]
F200W − F410M	0.901 [0.088, 1.009]	1.079 [1.020, 2.020]	13.000 [13.000, 21.000]	27.201 [24.332, 27.164]
F277W − F480M	0.482 [0.091, 0.589]	1.079 [1.020, 1.715]	16.770 [14.586, 20.548]	24.302 [22.429, 24.651]
F277W − F460M	0.592 [0.193, 0.705]	1.020 [1.020, 1.501]	13.000 [13.000, 20.984]	24.414 [22.771, 24.467]
F277W − F444W	0.239 [0.051, 0.485]	1.293 [1.020, 1.902]	16.863 [14.388, 17.279]	24.659 [22.949, 25.713]
F277W − F410M	0.097 [0.032, 0.376]	1.537 [1.020, 1.902]	16.480 [14.321, 17.083]	23.917 [22.461, 26.294]
F356W − F480M	0.102 [0.053, 0.303]	1.428 [1.020, 1.616]	16.348 [15.328, 20.632]	21.370 [21.000, 24.957]
F356W − F460M	0.246 [0.101, 0.344]	1.079 [1.020, 1.406]	19.807 [15.477, 20.062]	23.772 [21.862, 24.111]
F356W − F444W	0.070 [0.033, 0.180]	1.389 [1.020, 1.846]	15.631 [15.057, 18.192]	21.000 [21.000, 25.261]
F356W − F410M	0.120 [0.085, 0.127]	1.020 [1.020, 1.038]	15.277 [14.913, 16.393]	21.511 [21.000, 22.617]

Colour	SpT	Median (mag)	68.27% range (mag)	95.45% range (mag)
$F150W - F480M$	T6	0.337	$[-0.185, 0.859]$	$[-0.706, 1.381]$
$\Delta= 2.207$	T7	0.929	$[0.358, 1.501]$	$[-0.214, 2.073]$
	T8	1.750	$[1.122, 2.378]$	$[0.494, 3.005]$
	T9	2.820	$[2.130, 3.511]$	$[1.440, 4.201]$
	Y0	4.161	$[3.400, 4.922]$	$[2.640, 5.682]$
	Y1	5.791	$[4.952, 6.630]$	$[4.113, 7.469]$
	Y2	7.729	$[6.890, 8.569]$	$[6.050, 9.408]$
	Y3	9.992	$[9.153, 10.832]$	$[8.313, 11.671]$
	Y4	12.597	$[11.757, 13.436]$	$[10.918, 14.276]$
$F150W - F460M$	T6	0.132	$[-0.406, 0.671]$	$[-0.945, 1.209]$
$\Delta= 2.134$	T7	0.712	$[0.123, 1.300]$	$[-0.466, 1.889]$
	T8	1.526	$[0.881, 2.171]$	$[0.236, 2.816]$
	T9	2.579	$[1.871, 3.286]$	$[1.164, 3.993]$
	Y0	3.874	$[3.097, 4.651]$	$[2.320, 5.428]$
	Y1	5.416	$[4.561, 6.270]$	$[3.707, 7.125]$
	Y2	7.207	$[6.352, 8.062]$	$[5.497, 8.917]$
	Y3	9.251	$[8.395, 10.106]$	$[7.540, 10.961]$
	Y4	11.550	$[10.694, 12.405]$	$[9.839, 13.260]$
$F150W - F444W$	T6	0.745	$[0.352, 1.139]$	$[-0.041, 1.532]$
$\Delta= 2.008$	T7	1.224	$[0.797, 1.651]$	$[0.370, 2.078]$
	T8	1.936	$[1.471, 2.401]$	$[1.006, 2.866]$
	T9	2.895	$[2.387, 3.402]$	$[1.880, 3.910]$
	Y0	4.111	$[3.556, 4.667]$	$[3.000, 5.223]$
	Y1	5.596	$[4.986, 6.205]$	$[4.377, 6.815]$
	Y2	7.358	$[6.748, 7.968]$	$[6.138, 8.578]$
	Y3	9.405	$[8.795, 10.015]$	$[8.185, 10.625]$
	Y4	11.745	$[11.135, 12.355]$	$[10.525, 12.965]$
$F150W - F410M$	T6	1.075	$[0.732, 1.418]$	$[0.389, 1.761]$
$\Delta= 1.874$	T7	1.476	$[1.105, 1.847]$	$[0.734, 2.218]$
	T8	2.086	$[1.683, 2.488]$	$[1.281, 2.891]$
	T9	2.936	$[2.498, 3.375]$	$[2.059, 3.814]$
	Y0	4.059	$[3.580, 4.539]$	$[3.100, 5.019]$
	Y1	5.485	$[4.959, 6.012]$	$[4.433, 6.538]$
	Y2	7.244	$[6.717, 7.770]$	$[6.190, 8.297]$
	Y3	9.363	$[8.836, 9.889]$	$[8.309, 10.416]$
	Y4	11.871	$[11.344, 12.398]$	$[10.817, 12.924]$
$F200W - F480M$	T6	0.595	$[0.035, 1.155]$	$[-0.525, 1.715]$
$\Delta= 1.735$	T7	1.305	$[0.642, 1.969]$	$[-0.022, 2.632]$
	T8	2.210	$[1.418, 3.003]$	$[0.625, 3.796]$
	T9	3.305	$[2.349, 4.260]$	$[1.393, 5.216]$
	Y0	4.584	$[3.421, 5.746]$	$[2.259, 6.908]$
	Y1	6.044	$[4.620, 7.468]$	$[3.196, 8.892]$
	Y2	7.683	$[6.257, 9.109]$	$[4.831, 10.535]$
	Y3	9.499	$[8.073, 10.925]$	$[6.647, 12.351]$
	Y4	11.488	$[10.062, 12.914]$	$[8.636, 14.340]$
$F200W - F460M$	T6	0.376	$[-0.211, 0.962]$	$[-0.798, 1.549]$
$\Delta= 1.662$	T7	1.090	$[0.404, 1.776]$	$[-0.282, 2.462]$
	T8	2.018	$[1.211, 2.825]$	$[0.403, 3.633]$
	T9	3.131	$[2.173, 4.088]$	$[1.216, 5.046]$
	Y0	4.412	$[3.268, 5.555]$	$[2.124, 6.699]$
	Y1	5.849	$[4.474, 7.223]$	$[3.100, 8.598]$
	Y2	7.434	$[6.058, 8.810]$	$[4.681, 10.187]$
	Y3	9.160	$[7.784, 10.537]$	$[6.407, 11.913]$
	Y4	11.022	$[9.646, 12.398]$	$[8.269, 13.775]$
$F200W - F444W$	T6	0.912	$[0.480, 1.344]$	$[0.048, 1.776]$
$\Delta= 1.536$	T7	1.504	$[1.003, 2.005]$	$[0.502, 2.505]$
	T8	2.411	$[1.826, 2.997]$	$[1.241, 3.582]$
	T9	3.494	$[2.805, 4.183]$	$[2.116, 4.872]$
	Y0	4.673	$[3.856, 5.490]$	$[3.039, 6.307]$
	Y1	5.918	$[4.943, 6.894]$	$[3.967, 7.869]$
	Y2	7.216	$[6.239, 8.192]$	$[5.262, 9.169]$
	Y3	8.558	$[7.581, 9.534]$	$[6.604, 10.511]$
	Y4	9.939	$[8.962, 10.916]$	$[7.985, 11.892]$
$F200W - F410M$	T6	1.492	$[1.051, 1.933]$	$[0.610, 2.374]$
$\Delta= 1.402$	T7	1.829	$[1.323, 2.334]$	$[0.818, 2.839]$
	T8	2.510	$[1.928, 3.091]$	$[1.346, 3.673]$
	T9	3.491	$[2.817, 4.164]$	$[2.143, 4.838]$
	Y0	4.608	$[3.824, 5.392]$	$[3.040, 6.176]$
	Y1	5.780	$[4.863, 6.696]$	$[3.947, 7.612]$
	Y2	6.980	$[6.063, 7.897]$	$[5.145, 8.814]$
	Y3	8.200	$[7.282, 9.117]$	$[6.365, 10.034]$
	Y4	9.435	$[8.518, 10.352]$	$[7.601, 11.269]$
$F277W - F480M$	T6	1.400	$[1.072, 1.728]$	$[0.744, 2.056]$
$\Delta= 1.110$	T7	2.003	$[1.604, 2.402]$	$[1.204, 2.802]$
	T8	2.638	$[2.146, 3.130]$	$[1.654, 3.622]$
	T9	3.288	$[2.674, 3.903]$	$[2.060, 4.517]$
	Y0	3.950	$[3.174, 4.726]$	$[2.398, 5.502]$
	Y1	4.620	$[3.629, 5.611]$	$[2.638, 6.602]$
	Y2	5.297	$[4.304, 6.290]$	$[3.312, 7.282]$
	Y3	5.981	$[4.988, 6.973]$	$[3.995, 7.966]$
	Y4	6.670	$[5.677, 7.663]$	$[4.685, 8.655]$
$F277W - F460M$	T6	1.155	$[0.808, 1.502]$	$[0.461, 1.849]$
$\Delta= 1.037$	T7	1.776	$[1.365, 2.188]$	$[0.953, 2.600]$
	T8	2.423	$[1.929, 2.916]$	$[1.436, 3.410]$
	T9	3.077	$[2.479, 3.676]$	$[1.880, 4.274]$
	Y0	3.735	$[3.002, 4.469]$	$[2.268, 5.203]$
	Y1	4.396	$[3.488, 5.304]$	$[2.580, 6.212]$
	Y2	5.058	$[4.149, 5.968]$	$[3.240, 6.877]$
	Y3	5.723	$[4.813, 6.632]$	$[3.904, 7.541]$
	Y4	6.388	$[5.479, 7.298]$	$[4.570, 8.207]$
$F277W - F444W$	T6	1.775	$[1.475, 2.075]$	$[1.176, 2.375]$
$\Delta= 0.911$	T7	2.247	$[1.891, 2.603]$	$[1.535, 2.958]$
	T8	2.783	$[2.357, 3.209]$	$[1.930, 3.636]$
	T9	3.362	$[2.847, 3.878]$	$[2.331, 4.394]$
	Y0	3.975	$[3.346, 4.605]$	$[2.717, 5.234]$
	Y1	4.616	$[3.842, 5.390]$	$[3.068, 6.164]$
	Y2	5.282	$[4.506, 6.057]$	$[3.731, 6.832]$
	Y3	5.970	$[5.194, 6.745]$	$[4.419, 7.520]$
	Y4	6.678	$[5.902, 7.453]$	$[5.127, 8.228]$
$F277W - F410M$	T6	2.102	$[1.777, 2.428]$	$[1.451, 2.753]$
$\Delta= 0.777$	T7	2.474	$[2.106, 2.842]$	$[1.739, 3.210]$
	T8	2.897	$[2.480, 3.314]$	$[2.063, 3.732]$
	T9	3.363	$[2.888, 3.838]$	$[2.412, 4.314]$
	Y0	3.867	$[3.324, 4.411]$	$[2.780, 4.954]$
	Y1	4.406	$[3.783, 5.030]$	$[3.159, 5.654]$
	Y2	4.978	$[4.354, 5.602]$	$[3.729, 6.226]$
	Y3	5.580	$[4.956, 6.204]$	$[4.332, 6.829]$
	Y4	6.211	$[5.587, 6.836]$	$[4.963, 7.460]$
$F356W - F480M$	T6	0.417	$[0.054, 0.780]$	$[-0.309, 1.143]$
$\Delta= 0.621$	T7	0.733	$[0.305, 1.160]$	$[-0.122, 1.587]$
	T8	1.072	$[0.565, 1.578]$	$[0.058, 2.085]$
	T9	1.432	$[0.826, 2.037]$	$[0.221, 2.643]$
	Y0	1.811	$[1.082, 2.541]$	$[0.353, 3.270]$
	Y1	2.209	$[1.325, 3.093]$	$[0.441, 3.977]$
	Y2	2.624	$[1.739, 3.509]$	$[0.854, 4.394]$
	Y3	3.055	$[2.170, 3.940]$	$[1.285, 4.825]$
	Y4	3.501	$[2.616, 4.387]$	$[1.731, 5.272]$
$F356W - F460M$	T6	0.263	$[-0.129, 0.656]$	$[-0.522, 1.049]$
$\Delta= 0.548$	T7	0.617	$[0.170, 1.063]$	$[-0.276, 1.509]$
	T8	0.980	$[0.471, 1.490]$	$[-0.039, 2.000]$
	T9	1.351	$[0.765, 1.936]$	$[0.180, 2.522]$
	Y0	1.726	$[1.050, 2.401]$	$[0.374, 3.077]$
	Y1	2.105	$[1.321, 2.888]$	$[0.538, 3.671]$
	Y2	2.487	$[1.703, 3.271]$	$[0.919, 4.055]$
	Y3	2.872	$[2.088, 3.656]$	$[1.304, 4.440]$
	Y4	3.260	$[2.476, 4.044]$	$[1.692, 4.828]$
Colour	SpT	Median (mag)	68.27% range (mag)	95.45% range (mag)
$F356W - F444W$<br>$\Delta= 0.422$	T6	0.852	$[0.685, 1.019]$	$[0.518, 1.185]$
	T7	1.090	$[0.894, 1.285]$	$[0.698, 1.481]$
	T8	1.340	$[1.109, 1.572]$	$[0.878, 1.803]$
	T9	1.602	$[1.326, 1.878]$	$[1.051, 2.154]$
	Y0	1.875	$[1.543, 2.206]$	$[1.212, 2.538]$
	Y1	2.157	$[1.756, 2.558]$	$[1.355, 2.960]$
	Y2	2.449	$[2.047, 2.851]$	$[1.646, 3.252]$
	Y3	2.749	$[2.348, 3.151]$	$[1.946, 3.553]$
	Y4	3.058	$[2.656, 3.460]$	$[2.255, 3.861]$
$F356W - F410M$<br>$\Delta= 0.288$	T6	1.126	$[1.013, 1.239]$	$[0.900, 1.352]$
	T7	1.309	$[1.184, 1.433]$	$[1.060, 1.557]$
	T8	1.492	$[1.355, 1.629]$	$[1.218, 1.766]$
	T9	1.676	$[1.524, 1.827]$	$[1.373, 1.978]$
	Y0	1.860	$[1.692, 2.028]$	$[1.524, 2.195]$
	Y1	2.044	$[1.857, 2.231]$	$[1.671, 2.417]$
	Y2	2.229	$[2.042, 2.415]$	$[1.855, 2.602]$
	Y3	2.413	$[2.227, 2.600]$	$[2.040, 2.787]$
	Y4	2.598	$[2.412, 2.785]$	$[2.225, 2.972]$

Forward modelling results

Here we present the results of SED fitting with forward models. The fitting methods are described in § 5.4. Table B1 presents the results from fitting the NIRSpec/PRISM spectrum and MIRI photometry. Table B2 summarises the results from fitting the NIRCam and MIRI photometry, while Table B3 presents fits to the same photometry with additional priors derived from evolutionary models.

Data used for the colour-magnitude diagrams

Here we detail the data and references used to make the four colour-magnitude diagram panels in Figure 17.

Upper left: 2MASS $J - K_{\mathrm{s}}$ colour vs. 2MASS $J$ absolute magnitude. Comparison samples comprise M, L and T dwarfs from UltracoolSheet v2.1.0 ([22]) and the 20-pc census (Kirkpatrick et al. 2024) with parallaxes S/N $> 3$; direct-imaging M, L, T and untyped companions from the same catalogues; SUDA SPHEREx synthetic photometry for M–T dwarfs (Tu et al. 2026); JWST/NIRSpec late-T/Y dwarfs from Beiler et al. (2024b); MKO late-T/Y objects (Leggett et al. 2025) converted to 2MASS following Stephens & Leggett (2004); and the coldest dwarfs WISE 0855 (Luhman et al. 2024), WISE 1828 (Kirkpatrick et al. 2021) and WISE 1534 (Faherty et al. 2025).

Upper right: JWST/NIRCam $F150W - F444W$ vs. $F444W$ absolute magnitude. Data include SUDA L/T dwarfs; NIRSpec synthetic late-T/Y dwarfs (Beiler et al. 2024b); JADES T/Y candidates (Hainline et al. 2026b); NIRSpec L–Y dwarfs from RUBIES (Morrissey et al. 2026) and UNCOVER (Burgasser et al. 2024); brown dwarfs MoM-239450 (Hainline et al. 2026a), WISE 2220 (Albert et al. 2025), WISE 1828, and WISE 1534; Y-dwarf Capotauro (Liu et al. 2026a); and NIRSpec synthetic L-type planetary-mass companions TWA 27B (Luhman et al. 2023) and VHS 1256 b (Miles et al. 2023).

Bottom left: JWST/NIRCam $F150W - F480M$ vs. $F480M$ absolute magnitude. Data include SUDA M–T dwarfs; NIRSpec late-T/Y dwarfs from Beiler et al. (2024b); NIRCam/NIRISS late-T/Y dwarfs from Albert et al. (2025); an NIRSpec L–Y compilation from GTO 1189, GTO 1275, GO 3486, GO 3670, GO 4539, GO 5118 and GO 6084; metal-poor ultra-cool dwarfs from GO 4668; M–T dwarfs with SED-fitted distances from Tu et al. (2025a); direct-F480M Y-dwarf candidates from Hainline et al. (2026b); time-variable L/early-T objects WISE 1049AB (Biller et al. 2024; Chen et al. 2025b) and SIMP-0136 (Nasedkin et al. 2025); and two young M dwarfs from Manjavacas et al. (2024).

Bottom right: JWST/NIRCam $F200W - F444M$ vs. $F444M$ absolute magnitude. Data include SUDA, Beiler et al. (2024b), Tu et al. (2025a), Hainline et al. (2026b), the same NIRSpec program compilation, synthetic planetary-mass companions TWA 27B, VHS 1256b (Miles et al. 2023), WD 0806 b ([135]) and COCONUTS-2 b (Kiman et al. 2026b); direct-imaging companions 14 Her c ([12]), TWA 7 b (Lagrange et al. 2025; Crotts et al. 2025), AF Lep b (Franson et al. 2024), 51 Eri b (synthesized from its SPHERE spectrum; Brown-Sevilla et al. 2023; Madurouwicz et al. 2025) and RX J0534 b (Ferrer-Chavez et al. 2026); NIRSpec brown dwarfs in the Bullet cluster field (Bradac et al. 2026); and late-T planetary-mass companion Ross 458 C (Meynardie et al. 2025).

Photometry of other Y-dwarf candidates in JADES

This paper has been typeset from a TeX/LaTeX file prepared by the author.

Source ID	SpT	F070W	F090W	F115W	F150W	F162M	F182M	F187N	F200W	F210M
JADES-GN-BD-5†	Y0	–	$-0.52 \pm0.43$	$0.41 \pm0.34$	$-0.80 \pm0.48$	–	$0.44 \pm0.90$	$-5.13 \pm8.84$	$-0.41 \pm1.66$	$1.48 \pm1.58$
JADES-GN-BD-8	Y1(0.5)‡	$0.73 \pm0.87$	$-0.15 \pm0.51$	$1.11 \pm0.41$	$0.53 \pm0.52$	$0.35 \pm0.35$	–	–	$0.04 \pm0.44$	–
JADES-GN-BD-9†	Y1(0.5)‡	–	$-0.02 \pm0.35$	$0.92 \pm0.51$	$-0.22 \pm0.68$	–	–	–	$0.05 \pm0.44$	–
JADES-GN-BD-11	Y0	–	$-0.03 \pm0.28$	$0.34 \pm0.62$	$-0.41 \pm0.62$	–	–	–	$-0.06 \pm0.48$	–
JADES-GN-BD-16	Y1(0.5)‡	$1.89 \pm1.04$	$-0.17 \pm0.62$	$1.15 \pm0.66$	$-0.69 \pm0.63$	–	$-1.08 \pm1.33$	$1.03 \pm8.28$	$1.86 \pm0.79$	–
JADES-GS-BD-4	Y0	$-0.27 \pm0.44$	$0.13 \pm0.39$	$1.27 \pm0.36$	$1.29 \pm0.25$	$1.47 \pm0.52$	$0.50 \pm0.45$	–	$0.34 \pm0.46$	$0.70 \pm0.53$
JADES-GS-BD-8	Y1	–	$0.37 \pm0.27$	$3.77 \pm0.23$	$1.40 \pm0.25$	–	$0.31 \pm0.54$	–	$0.64 \pm0.31$	$1.63 \pm0.52$
JADES-GS-BD-21	Y1(0.5)‡	$-0.30 \pm0.90$	$0.15 \pm0.39$	$-0.44 \pm0.34$	$-0.01 \pm0.39$	–	–	–	$0.50 \pm0.33$	–
PANO-GS-BD-18	Y0	–	–	$14.16 \pm1.86$	$9.37 \pm1.51$	–	–	–	$4.04 \pm1.83$	–

Source ID	F250M	F277W	F300M	F335M	F356W	F410M	F430M	F444W	F460M	F480M
GNBD5†	–	$0.49 \pm0.66$	–	$1.75 \pm1.12$	$7.12 \pm0.61$	$37.56 \pm1.72$	–	$35.22 \pm1.60$	–	–
GNBD8	–	$2.08 \pm0.53$	–	$0.32 \pm1.65$	$12.28 \pm0.81$	$83.47 \pm3.07$	–	$95.01 \pm3.34$	–	–
GNBD9†	–	$0.75 \pm0.81$	–	$-0.28 \pm2.11$	$3.45 \pm0.93$	$24.90 \pm2.06$	–	$25.74 \pm1.34$	–	–
GNBD11	–	$-0.12 \pm0.78$	–	$0.88 \pm1.29$	$2.02 \pm0.81$	$11.45 \pm1.61$	–	$8.15 \pm0.87$	–	–
GNBD16	–	$0.80 \pm0.76$	–	$1.41 \pm1.86$	$1.74 \pm0.88$	$13.83 \pm1.31$	–	$17.95 \pm1.50$	–	–
GSBD4	$-0.14 \pm0.78$	$0.07 \pm0.52$	$1.09 \pm0.76$	$-0.96 \pm0.77$	$4.98 \pm1.75$	$24.84 \pm1.19$	–	$25.18 \pm1.08$	–	$22.57 \pm2.84$
GSBD8	–	$0.98 \pm0.28$	–	$-0.76 \pm0.61$	$7.74 \pm0.55$	$50.94 \pm1.68$	$67.82 \pm3.18$	$49.93 \pm1.63$	$38.29 \pm3.79$	$38.43 \pm2.59$
GSBD21	$-2.16 \pm3.21$	$0.30 \pm0.45$	–	$1.44 \pm1.42$	$1.37 \pm0.56$	$11.70 \pm1.35$	–	$15.78 \pm1.13$	–	–
GSBD18	–	$6.53 \pm1.50$	–	–	$38.98 \pm2.14$	–	–	$236.63 \pm7.48$	–	–