Anthropic's report on Claude scorpion toxin binders, published October 20, 2026, describes an internal research model that designed small engineered proteins against five scorpion neurotoxins. Seventeen of the 211 designs sent to the lab bound their intended targets. None has yet been shown to neutralize venom, which is the test that matters most for anyone who gets stung.
The project ran with the investment research firm Citrini Research, according to Anthropic's account. The company framed the work as a way to tackle diseases that attract weak commercial incentives. The model worked inside Claude Science, Anthropic's AI environment for researchers.
What the Claude model did on its own
Working largely autonomously, Claude used 12 protein design tools to generate 318,484 candidates and computationally screen 553,052 protein-toxin pairs, according to Interesting Engineering. The campaign ran for 64 hours and consumed about 10,000 GPU hours.
After an initial prompt and four follow-ups, the model selected 180 designs for laboratory testing. Researchers then added 31 more with the model's help, bringing the lab-tested total to 211.
That split shows where the work happened. The model handled most of the search and made the first cut, while people steered the project and added their own picks. The sources do not compare the time or cost of this approach with conventional screening, so the efficiency case rests on the raw counts alone.
How the lab tests decided which binders counted
The lab work was the real filter. Two outside contract labs, Adaptyv Bio and Twist Bioscience, tested the proteins with surface plasmon resonance, a method that measures how tightly molecules bind, according to TokenPost. Each design had to bind at both labs.
The bar was strict. A design had to bind its intended recombinant toxin, a lab-produced version, at both sites. It also needed a dissociation constant, or KD, of 1 micromolar or better, and it could not bind CLEC12A, an unrelated human protein, according to DEV Community.
Seventeen designs cleared every test. Here is how they split by toxin.
| Target toxin | Designs that passed |
|---|---|
| Cn2 (Mexican toxin) | 7 |
| AaH I | 4 |
| AaH II | 4 |
| LqhIII | 2 |
| Ts1 | 0 |
Which scorpion toxins the binders hit
Cn2 produced the most hits. The two strongest Cn2 binders reached single-digit nanomolar affinity and appear to sit at or near the site that an antibody neutralizing Cn2 in mice targets, according to TokenPost. That placement is a useful lead, but it is not evidence of neutralization.
No design bound more than one toxin. That means no single binder in this set covers several of the five targets, which is one reason the results are hard to read as a broad antivenom step.
Anthropic has a prior result to point to. In an earlier study, Claude designed binders against 14 of 15 target proteins, with hit rates of 22 to 35 percent compared with a typical 10 to 15 percent, The Decoder reported. By simple division, 17 of 211 is about 8 percent, below that earlier range. The sources do not say whether the two studies counted hits the same way, so the comparison is rough.
What the results do not show yet
The gaps are specific, and each one matters:
- Venom neutralization has not been tested. The team did not test whether any binder blocks venom.
- No cell or animal experiments have been reported.
- The tests used lab-produced toxins, not toxins taken from venom.
- Manufacturing cost and scale remain unresolved, according to RuntimeWire.
The accounts also disagree on autonomy. The Decoder says Claude worked alone for 24 hours after its initial prompt, while other reports describe a 64-hour campaign with researcher steering. Both descriptions may be accurate, but the sources do not give the exact split.
Independent reaction is thin. The Decoder paraphrases two outside experts who called the work a promising first step, but it does not provide their words verbatim, so this report does not quote them either.
Why scorpion venom research needs more than binding tests
The Decoder reports that scorpion neurotoxins kill more than 3,250 people a year. That toll is the reason the work matters to antivenom developers, but it does not make these binders a treatment.
A 2025 study in Nature offers a template. It reported designed proteins against snake toxins that neutralized them and protected mice, according to Kingy AI, which shows that designed binders can advance past binding tests. The Claude team has not reached that stage. Its scorpion binders remain at the binding step.
Binding is the first gate. Neutralization, animal protection and manufacturing are the gates that decide whether a protein could become part of a treatment. Anthropic's results clear the first gate for 17 designs and leave the rest open.
The pattern is not unique to Anthropic. Microsoft's MAI-DxO, which beat doctors on hard cases in a science study, is another example of AI taking on expert research work.
Policy questions follow. The sources I read did not address dual-use risk, meaning the chance that protein design tools could be misused. Policymakers and biosecurity researchers will need to weigh how autonomous design fits existing oversight. That concern is my inference, not a sourced finding.
How to read claims about AI-designed antivenom
Readers will see more headlines like this, so a few checks help:
- Identify what was measured. Binding, neutralization and animal survival are different results, and this study reports only the first.
- Read the affinity number. The bar here was a KD of 1 micromolar or better, which measures binding strength, not venom blocking.
- Ask whether the toxins were lab-made or venom-derived. These tests used lab-produced toxins.
- Look for outside testing. Adaptyv Bio and Twist Bioscience ran the assays here.
For consumers, there is nothing to act on yet. These binders are not a treatment or a product, and anyone who has been stung should keep following standard medical care. For a view of the company behind the work, see our report on Anthropic's IPO draft.
What comes next for AI-designed antivenom research
The next milestones are the ones the sources list as missing: neutralization tests against venom, cell and animal studies, and a plan for making the proteins at a workable cost and scale. Watch whether Anthropic publishes those results and whether any binder holds up against venom-derived toxins. Also watch for independent expert responses that go beyond paraphrase. This campaign suggests AI can cover a large search quickly. Whether that shortens the road to a usable antivenom is still open.
