Claude designs 17 protein binders against scorpion venom in 64 hours

Could a handful of financial analysts with zero drug development training design a working antivenom from scratch using an LLM?

Anthropic gave that question a real-world test, and the outcome is both genuinely impressive and a sharp reality check for AI pharma hype.

Anthropic published research showing how an internal version of Claude designed protein minibinders against scorpion neurotoxins. Working alongside five staffers from investment research firm Citrini Research, Claude generated over 318,000 designs in a 64-hour autonomous run. Subsequent wet-lab tests confirmed that 17 of the candidate proteins successfully bound to four deadly scorpion toxins.

A 35-page prompt and a $200,000 budget

Scorpion stings injure more than 1.2 million people annually and kill over 3,250, mostly children in underserved regions. Traditional antivenoms rely on milking venom and injecting it into horses to harvest antibodies. It is a slow, expensive process that rarely attracts pharmaceutical funding.

Citrini wanted to test whether nonexperts could tackle neglected diseases using Claude. After hitting Claude's biosecurity guardrails at their own desks, the Citrini team reached out to Anthropic. The AI lab agreed to lift those guardrails, invite them into its offices, and provide access to an unreleased internal model while footing a compute bill of just under $50,000.

Guided by a 35-page prompt and five human messages during the initial run, Claude deployed 12 design tools and soaked up roughly 10,000 GPU hours. The model selected 180 top candidates, added 31 more with human guidance, and directly emailed the sequences to contract research organizations.

Swiss biotech Adaptyv Bio and US-based Twist Bioscience synthesized and tested all 211 designs in physical labs. That testing phase cost around $150,000 and took nearly two months. Ultimately, 17 designs met binding criteria across four target toxins from Mexico, Brazil, North Africa, and the Middle East.

Binding is not a cure

Despite the buzz, Anthropic made sure to clarify that Claude did not invent a finished drug.

First, the team's bid for a universal antivenom failed entirely. The prompt pushed Claude to design multitoxin binders that could neutralize multiple venom variants at once. None of the candidates bound to more than a single toxin target.

Second, binding to a toxin in vitro does not mean neutralizing it inside a living body. Experts at the Technical University of Denmark pointed out that generating a molecule that clings to a toxin is now the comparatively easy part of the pipeline. Proving that it stops venom activity without causing toxicity, instability, or rapid degradation in an organism is the real hurdle. Anthropic has not tested whether these binders neutralize venom in cells or animals.

Anthropic released all 211 designs and their experimental data under a CC BY 4.0 license so other researchers can evaluate failures and test neutralization.

Why it matters

We think the most remarkable part of this experiment is not the biology, but who operated the keyboard. A team of financial newsletter analysts, possessing virtually no formal wet-lab drug development background, orchestrated a fleet of specialized computational biology tools and took a project from an abstract prompt all the way to physical laboratory validation.

However, this experiment cannot be taken as proof that anyone can now whip up treatments at home. Citrini received extraordinary access: white-glove onboarding inside Anthropic's headquarters, waived safety guardrails, expert staff supervision, and subsidized compute. For outside developers, Anthropic's Life Sciences Verification Program maintains strict vetting over who gets model access for biosecurity reasons.

Crucially, the experiment highlights where the true bottleneck in modern drug discovery lies. Claude cranked out hundreds of thousands of candidate designs in 64 hours for $50,000, but verifying just 211 of them took two months and $150,000. As generative models make molecular design cheap and ubiquitous, access to wet-lab bandwidth and animal testing will become the real constraint in biotech.