Claude wrote GPU code and made biology models 4x faster

Claude optimized inference for more than 30 open-source biology models and made them 4x faster on average, Anthropic said in a September 17 post on X. Part of the speedup came from custom GPU software that Claude wrote, and Anthropic is releasing all of the optimization code on GitHub.

Biologists use these models to work out the structure of molecular systems, design drug-like molecules and predict what genetic mutations do. Anthropic frames the work around cost: such models “are often expensive to run, potentially limiting their impact.” The details are in a Science Blog post, and the full results are in a technical report.

To show what the faster models make possible, Anthropic and Adaptyv Bio are running a protein design competition in which more than 5,000 designs will be tested in the lab. Anthropic is providing up to $1 million in Claude credits and is funding the lab testing together with Adaptyv. Modal is contributing up to $250,000 in compute, and Twist Bioscience is supplying the DNA.

Adaptyv’s Nipah Binder Competition, which ran from October 2025 to January 2026, tested 1,200 designs and had a 9.3% hit rate, according to Adaptyv’s Proteinbase page. The new competition is listed there as five challenges, each focused on a disease or a biological mechanism, running from September to December 2026.

All selected designs are made and tested in the same Adaptyv lab under the same protocols, and the data is published openly, including results for designs that did not work.