Claude's scorpion-toxin binders got their laboratory verdict on October 20, 2026. In a research report published that day, Anthropic said an internal research model, working inside Claude Science, ran largely on its own for 64 hours, generated 318,484 protein designs, and narrowed them to 211 candidates that Adaptyv Bio and Twist Bioscience then tested in the lab. Seventeen met the preset criteria for a binder, and all 211 designs and their data were released publicly.
Binding is not the same as neutralizing
Each of the 17 successful designs binds just one neurotoxin, covering four targets: Cn2 from Mexico, AaH I and AaH II from North Africa, and LqhIII from the Middle East. No candidate qualified against Brazil's Ts1, and every design meant to bind several toxins at once failed. The two strongest, Cn2-B1 and Cn2-B2, bound in the single-digit nanomolar range, and competition experiments suggest they attach at or near the site of an antibody reported to neutralize Cn2 in mice. Neither has been tested in cells or animals, and Anthropic is explicit that this is not an antivenom.
What the model decided during those 64 hours
The campaign ran from one initial prompt and four follow-ups. Claude picked the toxin structures from protein databases, marked five binding hotspots per toxin, set its own quality thresholds, and divided work across 12 design tools, screening 553,052 design-toxin pairs at a cost of roughly 10,000 GPU hours. Most shortlisted designs, including Cn2-B1, were selected only after the third follow-up, and Claude itself sent the sequences to Twist Bioscience for synthesis. Among single-toxin designs, 11 of 60 qualified, a hit rate of about 18%, in line with conventional protein-design campaigns.
Why scorpions, of all targets
Scorpions sting more than 1.2 million people a year and kill over 3,250, mostly children in places with limited medical care. Existing antivenoms are made from immunized horses over months, are expensive, cover few toxins, and are often in short supply, which is why scorpion envenomation is described as a neglected tropical disease. The project began when investment research firm Citrini Research asked whether motivated nonexperts with a model could take on a problem the market ignores. Two outside experts called the result a promising first step.
The bottleneck has now moved past binding. The candidates still need to be checked against toxins purified from real venom, tested with patch-clamp recordings to see whether they block the toxin's action on sodium channels, and then tried in animals, with safety, stability, and affordable production still ahead. The weight of the result is the compression: early discovery in days, lab confirmation in weeks, and the full dataset open for others to build on.
