Anthropic says a research model inside Claude Science produced 17 lab-confirmed toxin binders, and not an antivenom
Anthropic, Tuesday: an internal model, given one prompt and four follow-ups, ran for 64 hours inside Claude Science. It generated 318,484 protein designs. Adaptyv Bio and Twist Bioscience tested 211. Seventeen met Anthropic’s binding bar, each against a single toxin. Anthropic says it has not shown that any of them stops a toxin in cells or protects an animal, and that the toxins in the assays were made from the gene.
Feature4 minIn-depth
James Okonkwo4 min read
- Anthropic
- Claude Science
- Drug discovery
- Agents

Photo: AnthropicSource page
Anthropic said on Tuesday that an internal research model, working inside a system it calls Claude Science, designed small proteins that bind scorpion neurotoxins, and that outside labs confirmed 17 of them. The model worked from a prompt and four follow-ups, which Anthropic says it wrote with Citrini Research, a financial-investment research firm. The run took 64 hours. Anthropic’s own limit is in the same post: it has not shown that any design neutralizes venom in cells, and it does not claim an antivenom.
The count is the company’s. The model generated 318,484 designs and picked 180 for testing. Anthropic’s team added 31 more, with the model’s help. Adaptyv Bio and Twist Bioscience, the two contract labs, tested all 211. Seventeen met the bar Anthropic set, each against one toxin. Seven were against Cn2, four against AaH II, four against AaH I, and two against LqhIII. Anthropic says it also asked for binders against Ts1, and for binders that would hit more than one toxin. None of the multi-toxin designs met the bar against more than one. Of 60 designs aimed at a single toxin, 11 bound, which Anthropic calls 18 percent and says is in line with a typical protein-design campaign, including its own earlier one.
What the labs actually measured
Anthropic’s bar was binding, measured at the bench. A design had to show binding in every replicate, every measurement had to be 1 micromolar or tighter, and it had to show no signal against CLEC12A, an unrelated human protein Anthropic used as a check for stickiness. Both labs had to agree, except for LqhIII, which only Adaptyv tested. Adaptyv made the designs in a cell-free system. Twist made them in human kidney cells, fused to the stem of an antibody. The toxins in the assays were recombinant: grown from each toxin’s gene, not purified from venom. Anthropic says that is the first limit on the result.
The two strongest designs, which Anthropic calls Cn2-B1 and Cn2-B2, target Cn2, from the Mexican scorpion Centruroides noxius. Anthropic says both bound in the single-digit nanomolar range at both labs, tighter than the 1 micromolar bar. It also says competition experiments suggest they attach on or near the same site as an antibody called LR, which has been reported to neutralize Cn2 in mice. Anthropic then says it has not tested Cn2-B1 or Cn2-B2 in cells or in live animals. The suggestion and the missing test are both in the post. Anthropic is reporting where the designs bind. It says it has not tested whether they stop the toxin.
Mostly on its own, and then prompted again
Anthropic says the model worked mainly on its own, under supervision, and that the team aimed it with four follow-ups. It worked autonomously for 24 hours after the first prompt. The team then told it to keep going and to look for better designs. The follow-ups, Anthropic says, were mostly about a binder that would hit several toxins. The model picked most of the 180 designs, including Cn2-B1, only after the third follow-up. At the team’s request it emailed sequences to Twist, in messages Anthropic says were marked as written by the model. The campaign used on the order of 10,000 GPU hours. Anthropic notes that most researchers do not have that much, and says that separately it has made these models about four times faster on average, and that a simpler prompt, one agent and one GPU over 24 hours, reaches comparable scores on the computer. Those computer scores are not the lab result.
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. Anthropic says the neurotoxins most likely to harm people attach to sodium channels on nerve cells, and that it aimed the designs at each toxin’s channel-binding site. It gives the public-health figures it is arguing from: scorpions sting 1.2 million people or more a year, and more than 3,250 die, mostly children in places with limited medical care. Existing antivenoms, it writes, are generally made from antibodies of immunized animals, can take months, usually cover few toxins, and are expensive or scarce in many regions. The hope in the post is that a binder could one day become a treatment. The result in the post is the binding data.
Designing a binder against almost any target no longer seems unrealistic. We can now obtain very tight binders remarkably quickly, and binding itself is no longer the main bottleneck.
— Susana Vázquez Torres, Spanish National Cancer Research Centre, quoted by Anthropic, Oct 20, 2026
Torres, a postdoctoral researcher in Madrid, and Timothy P. Jenkins, an associate professor at the Technical University of Denmark, are the two outside experts Anthropic says it sent the work to. Jenkins called designing credible binders against five toxins in a matter of days a showcase, and said neglected diseases such as animal envenoming have had little commercial reason to attract the work. Torres’s full remark, as Anthropic prints it, puts the impact on everything after binding: molecules that can become drugs, and that are safe, affordable and widely accessible. Anthropic says the same thing in its own voice. It names Cn2-B1 as a candidate it would consider, and it says the study has not shown binding to toxin taken from venom, has not shown that any design stops a toxin from acting on a cell, and has not shown that any design protects an animal. A treatment, it adds, would also have to be safe, stable, affordable and available where stings happen. The release is the 211 designs and their binding data.


