Anthropic says Claude designed working protein binders against 14 of 15 targets

Anthropic has published research showing Claude can autonomously run substantial parts of a protein-design campaign, reporting that its models "designed protein binders against 15 targets, and succeeded against 14 of them." The post, published August 18, 2026, details work done through Claude Science, an environment that gives the model internet access, GPU compute, and lab-integration tools to run multi-day autonomous research sessions.

What's new

Anthropic ran the campaign using two Claude models — Mythos Preview and Opus 4.8 — operating with access to research papers, BioRxiv, Google Drive, Slack, Gmail, and GPU access to specialized protein-folding and design models. Claude worked in two modes: a multi-target mode running 48-hour sessions with up to 12,500 NVIDIA H100 GPU-hours, and a single-target mode running 24-hour sessions with up to 2,500 H100 hours per target.

The results: 354 confirmed binders across 14 of 15 targets, from 1,320 total designs generated. Success rates varied by configuration — Mythos Preview hit 26.7% in multi-target mode and 35.1% in single-target mode, while Opus 4.8 hit 22.6% in multi-target mode. Anthropic states the industry baseline for this kind of protein-binder design campaign is typically 10-15%, meaning Claude's best configuration performed roughly double to triple the norm. On at least six targets, Claude achieved high-affinity binders, and on four or more it matched or exceeded the best previously published results.

The same post covers a second application: Claude Opus 5 analyzing analytical chemistry data. Claude processed NMR spectroscopy data in 23 minutes and LC-MS data in 19 minutes, with hydrogen-count results matching lab analysis within 0.08 of a proton and a purity reading of 96.4% against a lab measurement of 96.33%.

Context

This extends a line of research Anthropic has been building out through Claude Science on using Claude for hands-on scientific work rather than just literature review or hypothesis generation — the model is running the design-analyze-iterate loop with real compute and real lab-adjacent tools, not just summarizing existing results. Anthropic frames this as incremental progress toward a stated longer-term goal: the company says it plans to "run the entire development process end-to-end across all drug modalities," extending beyond protein binders to the full range of drug development approaches.

Why it matters

A 22-35% success rate against a 10-15% industry baseline is a meaningful jump for a task — protein binder design — that traditionally requires specialized computational biology expertise and iterative wet-lab testing cycles. If the results replicate outside Anthropic's own reporting, it points to AI models taking on a more active, autonomous role in early-stage drug discovery rather than serving purely as an assistive tool for human researchers. The analytical chemistry results are a smaller but concrete demonstration of the same theme: Claude matching lab-grade instrument analysis in a fraction of the usual turnaround time. Anthropic's own framing — that "scientific judgment" is expected to keep improving as models improve — signals the company sees this as a growing product and research priority rather than a one-off demo.