Anthropic releases dataset behind Claude's 14/15 binder run

Claude designed protein binders that worked in wet-lab tests against 14 of 15 targets, and Anthropic has now posted the prompts, structures and campaign data behind those runs as a Hugging Face dataset under a CC-BY-4.0 license.

The accompanying research post, dated August 18, 2026, reports Mythos Preview and Opus 4.8 hitting overall bind rates of 26.7% and 22.6% across all targets, with Mythos Preview reaching 35.1% in a single-target setup. For context, the post states that "10 to 15% is typical in protein design campaigns today." External validation was run by Adaptyv Bio and Twist Bioscience, and the full campaign produced 354 confirmed binders from 1,320 total designs, using up to 12,500 NVIDIA H100 hours in multi-target mode. Against the RBX1 target, Mythos Preview posted a 40% hit rate versus 3.7% among competition participants.

The uploaded repository breaks out into a data/ directory holding the v1.0 release, a structure_and_pae/ folder with mmCIF files and Predicted Aligned Error data across multiple seeds, the prompt set that drove the runs, and a 33.2 MB documentation zip. Anthropic tags the dataset for surface plasmon resonance and biolayer interferometry, the two biophysical techniques used to confirm binding.

Anthropic itself frames the numbers as a floor rather than a ceiling: "We expect that in the hands of expert protein designers this approach would yield even stronger results, especially if they give Claude active guidance and feedback on intermediate results." Two researchers on our Who's Who list circulated the repo link the day it went up.