Anthropic has published results from a 64-hour autonomous campaign in which Claude generated 318,484 protein designs targeting five scorpion neurotoxins, with 211 candidates advanced to laboratory testing and 17 meeting the binding criteria, the largest published demonstration to date of an AI model driving an end-to-end protein design loop at scale. The study extends Anthropic's earlier protein-design work and offers a rare, fully disclosed look at how much of a modern biology workflow a frontier model can already run on its own.
How the Campaign Ran
During the 64-hour design campaign, Claude called on 12 protein design tools and ran structural predictions across more than 550,000 protein-toxin combinations. The first round ran continuously and autonomously for 24 hours, and the compute phase consumed roughly 10,000 GPU hours. Of the 211 candidate designs tested in the laboratory, 17 met the binding criteria against the scorpion neurotoxins. Two of the candidates targeting the toxin Cn2 reached dissociation constants in the single-digit nanomolar range, a level of binding affinity that biologists consider tight and specific.
What the Results Do and Do Not Show
Anthropic is careful about the limits of the result: the campaign has not demonstrated toxin neutralization, protection in animals, or an effective antivenom. Laboratory binding and a usable treatment remain separate milestones, and the company published all 211 candidate designs along with the experimental data so outside groups can attempt the next steps. That transparency is notable in a field where negative results often go unpublished, and it gives academic teams a concrete benchmark for evaluating their own design methods.
Why It Matters
The study lands as AI labs compete to show that frontier models can produce verified scientific artifacts rather than plausible-sounding text. A 64-hour, model-driven loop that moves from target selection through tens of thousands of designs to lab-confirmed binders, with the underlying data released for outside scrutiny, is a meaningful data point for AI-for-science timelines. It also illustrates the compute intensity of the approach: roughly 10,000 GPU hours for a single design campaign against five targets, a cost curve that will shape how widely such methods can be applied across basic research and drug discovery.
Designs generated: 318,484 in 64 hours
Lab-tested candidates: 211 Confirmed binders: 17
Structural predictions: more than 550,000 protein-toxin combinations
Compute consumed: about 10,000 GPU hours
Data released: all 211 candidate designs and experimental results
Sources: Anthropic, AI Impact Hub, Tencent Research Institute, Sohu, threerouter