Claude designs binding proteins on his own: 354 candidates confirmed in the laboratory out of 1,320 proposals

Anthropic launched its models on fifteen therapeutic targets and had the results tested in a wet laboratory. The success rates obtained are well above the usual range in the industry.

On August 18, 2026, Anthropic published the results of a campaign in which its models design de novo protein binders , proteins tailored to bind to a specific target — the basic building block of a large number of biotherapies. Claude Mythos Preview and Opus 4.8 worked on fifteen targets of clinical interest and produced 1,320 designs, 354 of which turned out to be real bonds once synthesized and tested.

The highlight is the success rate. Binders' design campaigns today hover around 10 to 15% success; the models reach 22.6% and 26.7% when they treat all the targets simultaneously in 48 hours, and up to 35.1% when working target by target. Fourteen of the fifteen targets yielded at least one validated candidate. The designs were synthesized and evaluated without retouching by Adaptyv Bio and Twist Bioscience, two contract laboratories outside of Anthropic.

A second part focused on analytical chemistry: Claude Opus 5 interpreted raw NMR and LC-MS data in 23 and 19 minutes, with conclusions aligned with those of the reference laboratory — 0.08 hydrogen difference on a count, 96.4% compared to 96.33% on a purity measurement. The bottleneck thus shifts to the experimental phase, which remains counted in weeks.