Anthropic says Claude designed 17 lab-confirmed proteins that grab scorpion neurotoxins, but nobody has shown they stop a sting yet

Anthropic said on Tuesday, Oct. 20, 2026, that an internal Claude research model, working inside Claude Science, designed small "minibinder" proteins against five scorpion neurotoxins, and that 17 of the 211 designs tested by two outside labs met its criteria for binding a toxin dangerous to humans.

The honest headline is in Anthropic's own fine print: these proteins stick to scorpion toxins in a test tube, and that's it so far. But the speed is the story. A model went from a prompt to 180 lab-ready candidates in under three days, on a disease that kills thousands of kids a year and that drugmakers ignore because it doesn't pay. If the hard part of drug discovery really is moving from "find a binder" to "make a cheap, safe drug," the AI labs just made the first step close to free. The second step is still slow, expensive and human, and nobody's agent is doing animal trials yet.

Anthropic asked an internal research model, running inside Claude Science, to design minibinders against scorpion neurotoxins. A minibinder is a small protein meant to attach tightly to a target. The prompt and four follow-ups were developed with Citrini Research, an investment research firm, which had asked whether AI could let motivated nonexperts tackle a real health problem. Anthropic did not name the internal model. Those lines are Anthropic's.

Over 64 hours, working mostly on its own, Claude used 12 design tools to generate 318,484 protein designs, co-folded and screened 553,052 design-toxin pairs, and picked 180 for testing. Co-folded means the computer predicted how each designed protein and its toxin would sit together in three dimensions. Anthropic's team added 31 more designs with Claude's help. The campaign used on the order of 10,000 GPU hours. A GPU hour is one hour on one of the graphics chips used to run models. Claude also chose the toxin structures, from the Protein Data Bank and the AlphaFold database, the spots on each toxin to aim at, and the quality thresholds. It analyzed the data, made the figures, and drafted the report under supervision. Those lines are Anthropic's.

The targets were five neurotoxins from four regions: Cn2 from Mexico, Ts1 from Brazil, AaH I and AaH II from North Africa, and LqhIII from the Middle East. The designs aimed at the part of each toxin that latches onto nerve sodium channels. A sodium channel is the gate a nerve cell uses to fire a signal. Those lines are Anthropic's.

Contract labs Adaptyv Bio and Twist Bioscience tested all 211 designs for binding, meaning whether the designed protein stuck to the toxin. Seventeen met Anthropic's criteria: seven against Cn2, four against AaH II, four against AaH I, two against LqhIII, and none against Ts1. Claude also tried designing binders that hit several toxins at once. None of those met the criteria for more than one toxin. Those counts are Anthropic's.

The limits are in Anthropic's own account. Binding was shown only against recombinant toxins, toxins made from the gene in a lab, not toxin purified from venom milked from a scorpion. No design was tested for neutralization in cells, or for whether it works in animals. Anthropic wrote, "We do not claim to have produced an antivenom." The post cites no peer-reviewed paper. Those limits are Anthropic's.

On why scorpions, the post says stings hit 1.2 million or more people a year and cause more than 3,250 deaths, mostly children in places with limited medical care. Today's antivenoms are made by immunizing horses with venom milked from thousands of scorpions. They cover only a few toxins, and they are expensive or scarce. Those figures are the post's.

Anthropic sent the work to outside experts. Susana Vázquez Torres, at CNIO in Madrid, Spain's national cancer research center, said "binding itself is no longer the main bottleneck." The challenge, she said, is molecules "that can become real drugs, ones that are safe, affordable, and widely accessible." A bottleneck is the slowest step. Timothy P. Jenkins, at the Technical University of Denmark, said designing credible binders against five toxins in days is "a great showcase of the AI-driven paradigm shift." Those quotations are theirs, as Anthropic prints them.

Anthropic is releasing all 211 designs, with sequences and lab results, on Hugging Face. In August it had shown Claude designing minibinders from a single prompt. Those lines are Anthropic's.

In plain terms, Anthropic said on Tuesday that an internal Claude model designed small proteins meant to grab five scorpion neurotoxins, two outside labs tested 211 of them, and 17 stuck to a toxin under Anthropic's own criteria. The proteins were tested against lab-made toxin. They were not tested in cells, in animals, or in people, and Anthropic says it has not made an antivenom. The designs and the lab data are public.