Claude scorpion toxin binders: Anthropic says 17 of 211 lab-tested designs met its criteria, but none is an antivenom

Anthropic says Claude designed candidate protein binders against scorpion toxins, molecules meant to attach to a toxin, in a research campaign it described on October 20, 2026. Two outside labs measured 211 of the candidates. Seventeen met Anthropic's stated criteria for a binder, each against a single toxin. Anthropic calls the work an early research result and says it does "not claim to have produced an antivenom." It is the latest case of Anthropic reporting Claude at work on a science problem, after its NIST curve-seed investigation.

This story is based on Anthropic's research post and the dataset it published on Hugging Face, checked October 23, 2026, with Anthropic's August 2026 protein-design post for context. Every result below is vendor-reported: Anthropic designed the study, set the criteria and reported the outcome, and the measurements were made by contract research organizations it hired.

The short version

• What Anthropic claims: Claude, running in Anthropic's Claude Science environment, produced candidate designs against five scorpion neurotoxins and shortlisted 180 for testing.
• Tested: 211 designs in total (180 from Claude, 31 added by the team), measured by two contract labs.
• Met the criteria: 17, each against one toxin. Anthropic reports 7 for the toxin Cn2, 4 for AaH I, 4 for AaH II, 2 for LqhIII and none for Ts1.
• Effort: 64 hours from the first prompt to the final list, including an initial 24 hours of autonomous work and about 10,000 GPU hours.
• The big limits: no design met the criteria for more than one toxin, and none has been tested for neutralizing venom in cells or animals.

What did Anthropic report, and who measured it?

Anthropic frames the project as a test of whether Claude could help motivated nonexperts make progress on a real health problem. It says the idea came from Citrini Research, which it describes as a financial investment research firm, and that the two adapted a prompt from an earlier Anthropic protein-design study. Scorpion stings are a neglected tropical health problem; Anthropic's post says scorpions sting more than 1.2 million people a year, causing more than 3,250 deaths, mostly among children, though it cites no source for those two figures.

The lab work was done by two contract research organizations, Adaptyv Bio and Twist Bioscience. Both measured binding using surface plasmon resonance (SPR), a standard lab method for detecting whether two molecules stick together and how tightly. Anthropic, not the labs, chose which designs to make and reported the overall result.

What counts as a "binder" here?

The number that matters most, 17, depends entirely on Anthropic's own definition. Anthropic says a design counted as a binder only if it:

• showed binding at both labs, in every replicate;
• had a dissociation constant (KD, a standard measure where a lower number means tighter binding) of 1 micromolar or tighter; and
• showed no signal against an unrelated human protein used as a "stickiness" control.

By that bar, Anthropic reports 17 of the 211 tested designs qualified. It says 11 of 60 single-target designs met the criteria, which it calls an 18% hit rate, and describes that as "consistent with typical protein design campaigns, including our previous work." The post gives no comparison figures for that claim. One toxin, LqhIII, was assessed at only one lab, so Anthropic notes it could not meet the two-lab rule as written; its two reported binders are therefore a partial exception to the stated criteria.

Last verified: October 23, 2026

How strong were the best results?

Anthropic reports that its two strongest designs bound Cn2, a toxin from the Mexican scorpion Centruroides noxius, with affinities in the single-digit nanomolar range at both labs. (Nanomolar is tighter than the 1-micromolar cutoff by roughly a thousandfold.) It says competition experiments suggest these two designs attach at or near the same site as an antibody that has been reported to neutralize Cn2 in mice. That is a hint about where they bind, not evidence that they neutralize anything. Anthropic's published dataset lists the per-design KD values measured by each lab.

What are the limits?

Anthropic is direct about what the work does not show, and these caveats are the story as much as the 17 binders:

• No neutralization testing. The designs were tested only for binding, not for blocking a toxin's effect, in cells or in animals.
• Lab-made toxins. Binding was measured against recombinant (lab-produced) toxins, not toxins purified from venom.
• No multitoxin success. None of the designs aimed at several toxins met the criteria for more than one. Anthropic offers possible reasons, including the prompt, the limits of its tools and whether such binding is biologically feasible.
• Not a drug. Anthropic says a treatment would also need to be safe, stable, affordable and available where stings happen, and that it has not produced an antivenom.

Two outside researchers Anthropic quotes strike the same note. Susana Vázquez Torres of the Spanish National Cancer Research Centre said the "real impact will come from tackling the parts of the therapeutic pipeline that go beyond binding." Timothy P. Jenkins of the Technical University of Denmark called it "a great showcase of the AI-driven paradigm shift" in therapeutic discovery. Anthropic says both viewed the work as a "promising first step." It published a technical report and released the 211 designs, their measurements and structures as a dataset under a CC BY 4.0 license, which lets outside scientists check the claims.

How does this fit Anthropic's earlier claims?

In an August 18, 2026 post, Anthropic reported that Claude designed binders for 14 of 15 target proteins, with hit rates it put at 22.6% to 35.1% depending on the model and setup, against what it called a typical 10% to 15% for the field. Those are also vendor-reported numbers. The scorpion work extends that line to a specific, hard public-health target and, unusually for a company post, foregrounds a failure (the multitoxin designs) and open data.

It also lands amid broader scrutiny of what AI agents actually accomplish in research. Epoch AI's InnovationEval found agents fell short of a human ML innovation and overstated their own results, and Sakana AI's Claude-based peer-review system caught planted errors but missed others. The lesson those stories share applies here: a vendor's framed result is a claim to check, not a settled finding. Anthropic's decision to publish the underlying data makes that checking possible. For how to read a lab's own safety and capability claims, see our guide to the Opus 5.5 system card.

What's confirmed, reported and uncertain

• Confirmed: Anthropic published the post, a technical report and a CC BY 4.0 dataset of 211 designs with lab measurements on October 20, 2026.
• Vendor-reported: the design counts, the 17 binders, the per-toxin breakdown, the affinities, the 64-hour and 10,000-GPU-hour figures, and the comparison with typical hit rates.
• Uncertain: whether any design neutralizes a toxin or works against venom; whether independent groups will reproduce the measurements; the two stated death and sting figures, which Anthropic doesn't source.