Anthropic's Claude Science builds first complete all-sky ultraviolet map, filling gaps with AI inpainting

Anthropic's Claude Science has produced what the company says is the first complete ultraviolet (UV) map of the entire sky. In tests of the gap-filling step, predictions deviated by about 10% on average from actual measurements.

Johns Hopkins astrophysicist Brice Ménard describes the project on the Anthropic website. The Decoder reported it on Oct 9, 2026.

What was built

Claude Science coordinates multiple AI agents. According to the account, those agents:

• downloaded data from several space missions
• calibrated the data
• merged the datasets into a single map

The ozone layer blocks UV light, so it can only be measured from space. That is why no full-sky UV map existed before. NASA's GALEX mission covered about two-thirds of the sky and skipped bright star-forming regions.

To cover the rest, the system used inpainting. In this technique, a model learns from existing data to reconstruct missing regions. In tests, its predictions averaged about 10% deviation from actual measurements. The source does not say how the tests were designed, which regions were held out, or whether the figure has been independently validated.

Why UV matters

UV light reveals dust lit up by starlight. Examples include clouds around young stars and rings left behind by stellar explosions.

What is not disclosed

The source material does not specify:

• which Claude model or models power Claude Science
• context window size
• pricing
• compute used or time taken
• benchmark scores
• whether the map or the pipeline is publicly available

The "first complete" claim comes from the Anthropic-hosted account. It is not an independently verified result. Filled regions are model reconstructions rather than direct observations.

ÉMénard says the map is meant to serve as teaching material. He suspects many scientists have been putting off similar projects that AI could now make feasible.

What this means

This is a demonstration of agentic data work, not a model release. The reported contribution is orchestration: pulling, calibrating and merging heterogeneous archival datasets, a task that is tedious but not conceptually novel.

The key caveat is the inpainted portion. A ~10% average deviation may be acceptable for a teaching resource and for broad structure. It is a limit for quantitative astrophysics, where reconstructed pixels should not be treated as measurements. Users will need clear per-pixel provenance flags separating observed from inferred data. The source does not say whether the map includes them.

The larger claim is about backlog. If agent systems can reliably handle multi-mission data wrangling, similar archival projects in other fields become cheaper to attempt. Whether Claude Science does this reliably across other domains is not yet shown by this single example.