Anthropic and Claude Science Build the First Complete Ultraviolet Map of the Sky

Anthropic has published the first complete map of the entire sky in ultraviolet. It was made by Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, together with Claude Science. The map combines far UV (154 nm) and near UV (232 nm). About a third of its area, including a significant part of the Milky Way plane, was not measured by telescopes but predicted by a model. Every pixel is labeled "measured" or "predicted" and comes with an uncertainty estimate.

Why there was no complete UV map

The ozone layer absorbs ultraviolet light, so it can only be observed from space. Over half a century, telescopes have accumulated a lot of data, but the sky is covered unevenly. The largest dataset came from NASA's GALEX mission (2003–2013): about 38 000 individual observations covering roughly two thirds of the sky. However, GALEX deliberately avoided regions with very bright stars to keep from damaging its detectors. Those are exactly the regions that lie in the Galactic plane, where stars are densest. Swift and the South Korean FIMS/SPEAR added data, but the combined dataset still had many gaps.

Statistical methods for estimating the missing regions are known. The problem is the labor: they require pixel-level calibration and many repeated analyses, which means weeks of painstaking work. Ménard admits that astrophysicists put such tasks off in favor of more urgent ones. As a result, he had to apologize in class for a UV map full of holes. Every field has the same kind of backlog: projects that would explain a key idea or help colleagues, but never reach the top of the list.

How the agents built the map

In summer 2026, Ménard gave Claude Science a task that is simple to state and hard to carry out: collect all available UV surveys, bring them to a common scale, merge them into a single map, and fill in the regions that no UV telescope has seen. Claude acted as the orchestrator of a team of AI agents, and the work proceeded in several stages:

• agents searched the web for public astronomical UV surveys, each of which contains thousands or tens of thousands of images;
• the downloaded data was made internally consistent so that frames taken in different years could be compared;
• glare was removed from areas around bright stars, like the flare from a streetlight in a photo, since faint UV emission nearby cannot be measured otherwise;
• different regions of the sky were processed in parallel, each by its own agent or group of agents;
• the surveys were calibrated against one another, resampled to a single resolution and mapped to a common coordinate system.

Ménard gave the high-level instructions for merging the data, and Claude handled the details. Further specifics, including the method for predicting the missing regions, are not disclosed in the portion of the material available. All that is known is that it is described in Anthropic's original publication.

Where the result will be useful

The first use is educational. Students will be able to see the structure of the Milky Way in UV: clouds around young stars, rings left by supernovae, and dust lit up by starlight. The "measured/predicted" layers and uncertainty maps make it possible to separate real data from modeled data. For scientific work this is essential: predicted regions cannot be treated as observations, but they can be used with the estimated error taken into account.

The main takeaway reaches beyond astronomy. A "someday" task, for which a specialist did not have weeks to spare, was solved by coordinated agents in a single summer project. For research groups, this is a signal to review their backlogs for work with a clear problem statement, open data and cumbersome but formalizable calibration. Agent systems take on exactly such tasks most readily. And the uncertainty layer shows how to label AI output so that it can be verified.