Anthropic's Claude designs lab-confirmed protein binders

Anthropic published results on August 18, 2026 from a protein-design campaign run by Claude with almost no human steering after the opening prompt. Two contract labs, Adaptyv Bio and Twist Bioscience, independently synthesized the designs and measured whether they bound. Of 1,320 designs against 15 targets, 354 bound, a hit rate of about 27%. Anthropic says 10 to 15% is typical for this kind of campaign.

A minibinder is a small protein built to grab a specific target. Many medicines work that way: they attach and then block, activate, or deliver something. Designing one from scratch has historically taken a specialist weeks or months per target. Specialist machine-learning tools already speed up the computation. The slow part is choosing where to bind, running those tools, iterating, and waiting on the wet lab.

What did Claude actually do?

Anthropic gave Claude Opus 4.8 and Mythos Preview a long protocol, internet access, papers, and GPUs, then left the sessions alone except to approve network requests. Claude chose where on each target to bind, ran open-source structure, sequence, and co-folding models, and delivered 30 ranked designs per target. Campaigns ran 24 to 48 hours of wall time. The models it used are public. Anthropic is releasing the prompts, the computational models, and the binding data.

John Jumper, who co-created AlphaFold and shared the 2024 Nobel Prize in Chemistry, left Google DeepMind for Anthropic in June. This result is not a new folding model. It is an agent running tools the field already has, then handing the output to a wet lab.

How well did it work?

In a 48-hour session against all targets at once, Mythos Preview's hit rate was 26.7% and Opus 4.8's was 22.6%. When Mythos Preview worked one target at a time in 24-hour sessions, the hit rate rose to 35.1%. Anthropic says some of the strongest designs bound several times more tightly than the best previously published result for those targets. On RBX1, from an open design competition, Mythos Preview's single-target run hit 40% against 3.7% among the contest entries, and its top design beat the winner when both were re-tested on the same plate. Against maltose-binding protein, none of 90 designs bound.

The tightness comparisons and the "typical 10 to 15%" figure are Anthropic's. The binding itself was measured by the outside labs.

Can you use this in Claude?

Not the protein-design loop. Anthropic says life-science research of this kind is blocked in its most capable generally available model, and that protein design and other dual-use biology stay off Fable 5 while it builds a trusted access program for scientists. Mythos 5 has been a restricted model since the US cleared it for trusted partners in June.

The chemistry result is closer to the product. Claude Opus 5, given a contract lab's raw NMR and LC-MS files and a two-sentence prompt, returned processed results in 23 and 19 minutes and matched the lab on hydrogen counts and on purity, 96.4% versus 96.33%. NMR checks they made the molecule they meant to. LC-MS checks purity. Anthropic says that task can be tried in Claude Science with a raw file.

What is still unknown

A confirmed binder is not a drug. Anthropic said a scientist access program is a priority and that it expects to share more soon, without giving a date. Whether the hit rates hold once other groups rerun the prompts is untested.