Anthropic: Claude finds ART system in viral DNA

Anthropic frames ART as the first result from its biology lab, and the path there was a search rather than a single flash of insight. Roughly 950 agents worked for 21 hours, gathering records of more than 200,000 reverse transcriptases and narrowing about 3,500 candidates down to roughly 20 detailed reports. The enzyme at the center of ART was already on record; what Claude drew attention to were the parts around it, including the repeating DNA that resembles part of CRISPR.

The division of labor is notable. Scientists picked the research area and performed all the physical lab work, while Dario Amodei described the effort as "mostly, though not entirely" Claude's. Those human experiments produced a concrete lead — distinct short RNA molecules made from the repeat array — but not an answer about what ART does or whether it can cut, copy and paste DNA.

Anthropic places the finding alongside OpenAI's recent claims in mathematics, including a September 21 statement that an internal model resolved more than 100 longstanding open problems, and an earlier claim of a solution to the Navier–Stokes Millennium Prize problem. Whether agent-driven discovery becomes routine in biology appears to hinge on how many leads survive testing and how much experimental capacity labs can devote to checking them.