Anthropic says Claude Science has produced the first complete map of the sky in ultraviolet (UV) light. In a post published on October 8, 2026, Anthropic describes how astrophysicist Brice Ménard of Johns Hopkins University, who is also a researcher at Anthropic, had Claude orchestrate a team of agents that merged every public UV survey it could find and then predicted the roughly one third of the sky that no UV telescope has ever observed. The predicted regions matched hidden test patches to within about 10 percent. It is a clean example of an agent pipeline doing a long, tedious scientific data-engineering job, and a good case study in what to trust and what to check.
Quick facts
| Item | What Anthropic reports |
|---|---|
| Who | Brice Ménard (Johns Hopkins University, Anthropic researcher) with Claude Science |
| Output | First complete all-sky map in UV (far-UV 154 nm and near-UV 232 nm combined) |
| Observed vs predicted | About two thirds of the sky had UV data; roughly a third was predicted |
| Main data source | NASA GALEX (2003 to 2013, about 38,000 observations) plus Swift, FIMS/SPEAR, TD-1 |
| Extra inputs | Visible, infrared and radio maps (including Planck), and Gaia star data for more than 100 million stars |
| Self-test accuracy | About 10 percent on deliberately hidden regions |
| Effort | More than a dozen versions over several days |
Why there was no complete UV map
UV light is absorbed by Earth's ozone layer, so the only way to see the sky in UV is from space. According to Ménard, space telescopes have covered parts of the sky over 50 years, but the largest dataset, NASA's GALEX mission, imaged about two thirds of the sky and deliberately skipped locations with very bright stars, including much of the Milky Way's plane. That is exactly where dust glowing in starlight is most interesting, because very bright UV sources risked damaging the satellite's detectors. Swift and South Korea's FIMS/SPEAR added more coverage, but gaps remained.
Statistical methods to estimate the missing data exist, but Ménard says doing it properly takes weeks of pixel-level calibration and repeated analysis. His point is that most scientific fields have a backlog of such projects: useful maps, figures and reference resources that never rise to the top of anyone's list. That framing matters more than the map itself. The claim is not that Claude made a discovery; it is that Claude made a low-priority but valuable project cheap enough to actually finish.
How the pipeline worked
Ménard's instruction was simple to state: gather every available UV dataset, put them on a common scale, merge them, and fill every unobserved patch. Claude Science then ran a team of agents through several stages.
- Find the data. Agents searched the web for public UV surveys, each containing huge collections of images or measurements taken over years under varying conditions.
- Make each survey internally consistent. Images from different periods had to be comparable. Special care went to regions near bright stars, where glare has to be removed before the faint UV light nearby can be measured. Many agents worked in parallel on different regions of sky.
- Cross-calibrate and merge. Because each instrument sees the UV sky slightly differently, the surveys were cross-calibrated, redrawn at the same resolution, and mapped to a common coordinate system.
- Fill the gaps. Claude used inpainting, a standard machine learning technique where a model learns how parts of an image relate to their surroundings and restores missing regions. It combined that with other wavelengths: from the two thirds of the sky mapped in UV, it learned how UV brightness relates to visible, infrared and radio data, then applied that to the unobserved third, with an estimate of its own confidence at each point.
- Add the stars. On top of the diffuse inpainted background, Claude added UV estimates for more than 100 million individual stars, inferred from visible-light measurements by ESA's Gaia satellite.
The workflow rhythm, as Ménard describes it, was a few planning exchanges inside Claude Science, then hours of independent computation by the agents while he worked on other projects.
The checks, and the honest limits
The most important detail is how the result was validated. Ménard had Claude take regions where real UV data exists, hide parts of them, and reconstruct the hidden pixels. After several refinement rounds the estimates landed within about 10 percent of the real measurements, which Anthropic calls almost imperceptible to the eye.
Read that carefully. It shows the method works on sky that resembles the training sky. It cannot prove the predictions are equally good on the regions that were never observed, which are systematically different: they are mostly the bright-star and galactic-plane areas GALEX avoided. That is why the published map carries extra layers labeling each pixel as measured or predicted and giving an uncertainty estimate. Anyone using the map for research should work with those layers rather than treat predicted pixels as observations. The post is also a first-party account written by someone employed by Anthropic; it is not a peer-reviewed paper, and we have not seen independent replication.
Where the agents went wrong
The post is useful because it admits failure. Looking through processed images one evening, Ménard noticed faint circles in one of the dimmest fields, each slightly brighter or darker than its neighbors. They were the footprints of individual GALEX observations: each image captures a circular patch and contains a slightly uneven UV glow from Earth's atmosphere, and if that glow is not fully removed each circle stands out. Claude had listed this as a known issue at the start, yet the map still passed two rounds of review by other agents without the problem being caught.
When Ménard told Claude he could see discs from individual observations, the agents traced the cause to leftover atmospheric glow and corrected it across all 38,000 observations in a couple of hours. Two lessons follow. Agent review loops can share blind spots with the agents that produced the work, and a domain expert looking at the output still catches things automated review misses. Ménard's contribution, in his words, was guiding the process, and the expert-in-the-loop part was not decorative.
Why this matters beyond astronomy
This sits in a growing line of Anthropic science work. We covered the launch of Claude Science as an AI workbench for scientists, the discovery of a novel enzyme system with CRISPR-like repeats, the Yang-Mills amplitude result, and the company's physical wet lab. The UV map is different in kind: it is not a headline discovery but a piece of scientific infrastructure, a reference dataset that a researcher would otherwise have postponed indefinitely.
Other labs are working the same seam. Google's fruit-fly brain connectome work and the Novo Nordisk use of Claude Science in drug discovery both show AI absorbing the grinding data-processing work around science. Compared with those, the astronomy case has an unusually clear structure: public data in, a reproducible pipeline, a self-test, and an output anyone can inspect visually.
What you can take from the method
- Treat predicted data as a labeled layer. Keeping measured and predicted pixels separable, with uncertainty, is the pattern that makes AI-filled data usable in science.
- Hold out data to test. Hiding known regions and reconstructing them is a cheap, general validation for any gap-filling job, though it measures in-distribution accuracy only.
- Use other modalities as side information. The gap-filling used other wavelengths where UV was missing. The same idea applies whenever one channel is incomplete but correlated channels exist.
- Keep a human reviewing the output. The circles artifact was caught by eye, after agent review missed it.
- Parallelize by region. Splitting sky regions across many agents is the kind of embarrassingly parallel decomposition that agent systems handle well. For more on running agent teams, see our guide to multi-agent graph engineering.
What to watch next
Three things would turn this from a good demo into a lasting resource: an independent comparison of the predicted regions against future UV observations, publication of the pipeline details so others can reproduce or improve it, and adoption of the measured-versus-predicted layers by astronomers or educators. Ménard says the map is meant as a teaching tool, letting students see the structure of the Milky Way in UV alongside the familiar visible, infrared and radio views. The map is a legitimate use case for AI today; whether it stands up as a research dataset depends on follow-up checks that nobody outside Anthropic has run yet.
Check back on explainx.ai and follow @explainx_ai for updates as outside astronomers respond.
Related reading
- Claude Science: Anthropic's AI workbench for scientists
- Claude discovers a novel enzyme system with CRISPR-like repeats
- Claude and nine-loop Yang-Mills amplitudes
- Anthropic's physical wet lab
- Novo Nordisk and Claude Science in drug discovery
- Google's fruit-fly brain connectome with AI
- Official: Anthropic, The missing map of the sky