Claude designed working protein binders for 14 of 15 targets

Anthropic says Claude designed working protein binders against 14 of 15 targets in a wet-lab-tested campaign, an early sign that general AI agents can help with parts of experimental life-science research.

Protein binders are small proteins built to attach to specific targets. They are useful in drug discovery because many medicines work by binding to a biological target and changing its behavior. Anthropic says Claude Opus 4.8 and Mythos Preview generated 1,320 designs, with outside labs Adaptyv Bio and Twist Bioscience validating 354 binders across 14 targets.

The reported hit rates were 22.6% for Opus 4.8 and 26.7% for Mythos Preview when Claude worked across all targets in one 48-hour session. In a single-target setup, Mythos Preview reached 35.1%. Anthropic says 10-15% is typical for protein-design campaigns today, and that some Claude designs matched or exceeded the best reported affinity for several benchmark targets.

Anthropic also tested Claude Opus 5 on analytical chemistry. Given raw NMR and LC-MS files plus a short prompt, the model produced finished analyses in 23 and 19 minutes, matching a contract lab on hydrogen counts and near-identical purity values.

The caveat is important: these are company-reported results, not a finished drug-discovery system. Anthropic also says life-science tasks are blocked in its most capable model for now, with a scientist access program planned. Still, this is more than a benchmark: the output was physically tested, which makes the result worth watching.