Claude Designed 17 Scorpion-Toxin Binders: What the Result Does and Does Not Show

Anthropic says an internal Claude research model ran most of a protein-design campaign aimed at scorpion neurotoxins, and that 17 of 211 designs sent to outside labs met its binding criteria. The company is explicit that it did not show venom neutralization in cells or animals, so this is not an antivenom. It is an early, narrow and interesting result.

What did Anthropic actually report?

Anthropic reported that Claude designed small proteins, called minibinders, intended to latch onto scorpion neurotoxins, and that a subset of them did. According to Anthropic's write-up, the model generated around 318,000 candidate designs and picked 180 for testing; researchers added 31 more, for 211 total. Independent contract labs tested all of them, and 17 cleared the binding threshold.

The division of labor matters as much as the number:

• Human researchers wrote the opening prompt, sent four follow-ups, and supervised throughout.
• Claude chose which toxin structures to target, set the quality thresholds, analyzed results, produced figures, and drafted the report.
• Adaptyv Bio and Twist Bioscience handled the physical lab work, so the validation happened outside Anthropic.
• Citrini Research, an investment research firm, helped develop the prompts, which is an unusual collaborator for a bioscience paper.

Anthropic also released data for all 211 designs, including the failures, which is the part working scientists tend to care about most. Two outside researchers, Susana Vázquez Torres and Timothy P. Jenkins, reviewed the work and described it as a promising first step. Benzinga's account refers to five target toxins, while Anthropic's page describes the confirmed binders as each hitting one of four. Small discrepancy, but the kind of detail a headline flattens.

Why does scorpion antivenom need new approaches?

Because the current supply model is species-specific, fragile, and often missing where it is needed. A 2025 opinion article in Frontiers in Public Health cites estimates of more than 1.2 million scorpion stings and over 3,000 deaths a year, with children carrying a disproportionate share of the burden and mortality concentrated in low-resource settings.

The same article describes antivenoms as highly species-specific with limited cross-neutralization, often weaker outside the native range of the species they were made for, and produced in many low- and middle-income countries on limited capacity and aging infrastructure. The World Health Organization describes a similar picture for snakebite envenoming, estimating 81,410 to 137,880 deaths annually and warning that supply failure may be imminent in parts of Africa and Asia.

So the need is real, and a cheaper, more storable alternative to animal-derived antivenom would matter, independent of who or what did the designing.

What is the difference between binding and neutralizing?

Binding means a designed protein sticks to a toxin. Neutralizing means it stops the toxin from harming a living system. The first is a laboratory measurement; the second is a biological outcome, and the gap between them is where most candidate therapies die.

Anthropic drew that line itself. Its stated limitations include that the binders were tested against lab-made toxins rather than material taken from venom, that it has not confirmed neutralization in a cellular environment, and that none of the designs aimed at several toxins at once met the criteria. A single-toxin binder is less useful in the field, where a clinician often does not know which scorpion was involved.

Has AI-designed antivenom work been done before?

Yes, and further along. In January 2025, the Institute for Protein Design at the University of Washington announced work published in Nature on designed proteins that bind three-finger toxins from elapid snakes such as cobras and mambas. That study, described by the IPD and published as "De novo designed proteins neutralize lethal snake venom toxins", reported neutralization in laboratory tests and protection in mice, with survival rates reported in the 80-100% range depending on the pairing.

That is a useful yardstick. The snake toxin work reached animal protection; the scorpion work reached binding. Vázquez Torres, who led the earlier study, is one of the outside reviewers on the Anthropic work, so the comparison is not being hidden.

What do skeptics say?

Mostly that the autonomy framing outruns the evidence. The recurring objections:

1. Volume is not insight. Peter Clark of the Allen Institute for AI told Science News that separating junk hypotheses from valid ones is a "big, big problem." Generating 318,000 designs is cheap; the filtering is the science.
2. Hype is doing marketing work. Gary Marcus has argued that the transformation of science by language models "is not really happening yet."
3. Prediction is not understanding. Harvard's Keyon Vafa found a model that predicted planetary orbits without learning gravity.
4. The humans are load-bearing. A prompt, four follow-ups, chosen tooling, and two contract labs is a substantial scaffold. Those competing views are laid out in Science News, alongside researchers who think real discoveries are already happening.

A 17-for-211 hit rate is also hard to grade without a human baseline run the same way, and Anthropic did not provide one.

How should brands communicate an AI breakthrough claim?

Carefully, because the cost of overstating is asymmetric: a claim that gets walked back damages trust far more than a modest claim helps. The Anthropic post is a decent model here, since it published its own limitations and its failed designs. Most marketing teams do neither.

A short discipline for any AI result you plan to promote:

• Name the measured endpoint, not the hoped-for one. "Binds in vitro" is not "treats stings."
• State the human contribution plainly. Prompts, supervision, and third-party validation are part of the claim.
• Publish the denominator. 17 looks different next to 211.
• Say what you did not test. Readers who find the gap themselves assume you hid it.

This is where AI visibility connects to brand trust. Generative engines now summarize your claims for people who never reach your page, and an overstated claim gets compressed into something worse. MarkGrid's Model Share tracks how your brand and its claims are represented across AI models, and our GEO guide covers writing so summaries stay accurate. See also our overviews of AI brand intelligence and AI visibility metrics worth tracking.

FAQ

Did Claude create an antivenom?

No, and Anthropic does not claim it did. The designs were shown to bind toxins in lab conditions, and the company says it has not confirmed neutralization in a cellular environment or in animals.

How autonomous was the campaign?

Partly. Anthropic says Claude worked largely on its own after an initial prompt and four follow-ups, choosing targets and thresholds and drafting the report. Humans wrote the prompts and supervised, and outside labs did the physical testing.

Why does it matter that each binder hit only one toxin?

Because real stings do not come labeled. Antivenoms covering multiple related toxins are more practical clinically, and Anthropic reported that none of its multi-toxin designs met the criteria.

Is 17 out of 211 a good result?

It is a real result, but hard to rank without a matched human-led comparison, which was not published. The more informative signal may be that Anthropic released the full dataset including failures.

How does this compare to the snake venom work?

The 2025 Nature study from the Institute for Protein Design reported neutralization in lab tests and protection in mice. The scorpion work stopped at binding, an earlier stage of the same pipeline.

What is the practical lesson for communicators?

Match the claim to the endpoint actually measured, disclose the human role, give the denominator, and name the untested parts. Those four habits cost nothing and prevent most credibility damage.

Last updated October 22, 2026.