Anthropic says its Claude models just found a new gene-editing enzyme system on their own, with no scientist steering the search.
The finding comes from Anthropic's newly launched wet lab, its first attempt to test whether Claude can do real biology work rather than just answer questions about it. Nearly 950 Claude agents spent 21 hours combing through a massive database of DNA sequences, burning through about 210 million tokens before one agent flagged an unusual repeating pattern. Anthropic says that pattern pointed to an enzyme system similar to the molecular machinery behind Crispr, the gene-editing tool that reshaped biology research. The company says its scientists only wrote the initial prompt and ran the physical lab work needed to confirm what the model found.
This is Anthropic's biggest science claim yet, and it lands right as the company prepares to go public and needs proof its models are worth more than a very good chatbot. A self-reported, in-house discovery is not the same as a validated tool: Crispr itself took years of independent confirmation before labs trusted it, and Anthropic hasn't said whether this result has been peer reviewed or published anywhere.
DeepMind spent years racking up benchmarked, peer-reviewed wins with AlphaFold before anyone called it a breakthrough - Anthropic is asking us to take its word for this one on day one.