What happened
Anthropic published research on August 18 showing that Claude Opus 4.8 and Mythos Preview autonomously designed protein binders—small molecules that latch onto disease targets—a critical early step in drug discovery. Given 15 biological targets, the models succeeded on 14, generating 1,320 candidate designs of which 354 were experimentally validated as functional by two independent labs, Adaptyv Bio and Twist Bioscience.
The numbers
Claude's success rate ranged from 22% to 35% depending on mode and model, more than double the typical 10-15% industry baseline. Against RBX1, a notoriously difficult target, Mythos Preview hit 40% compared to 3.7% from human entries in a public competition. The models ran for 24-48 hours with up to 12,500 Nvidia H100 GPU hours, using only a single expert-written prompt, internet access, and open-source tools. In a separate test, Opus 5 analyzed raw lab instrument files in 19 minutes and measured sample purity at 96.4%—matching the lab's own 96.33% result that took four days to produce.
Why it matters
CEO Dario Amodei said days earlier that biology breakthroughs were "months away." This result arrived in days. The difference: this is a general model running an entire workflow autonomously, not a narrow domain tool. You're not building on your stack; you're orchestrating Anthropic's reasoning as a service, which now quietly controls your protein-design pipeline. When intelligence becomes infrastructure, owning the data and owning the outcome split apart.