Anthropic's Claude AI Generates Hundreds of Valid Protein Binders Across Most Test Targets

Anthropic reported that its Claude AI suite designed working protein binders for 14 of 15 test proteins, a result validated by external laboratories that synthesized and screened the candidates. Out of 1,320 proposed sequences, 354 bound successfully, giving hit rates of 26.7% for Mythos Preview and 22.6% for Opus 4.8 when tackling all targets simultaneously, rising to 35.1% when Mythos Preview focused on one target at a time—well above the typical 10-15% success range.

The AI operated with 48 hours of wall time, up to 12,500 NVIDIA H100 GPU hours, and a 30,000-token prompt, handling design, folding, and optimization while humans only approved access and ordered the final hits. In a separate test, the publicly available Opus 5 parsed raw NMR and LC-MS files and delivered purity results within 23 and 19 minutes, matching a lab’s four-day report. Anthropic noted mixed outcomes on specific proteins such as RBX1, TNFα, and maltose-binding protein, and acknowledged the need for a scientist-access program for its most capable model, Fable 5.