On October 20, 2026, Anthropic published a report on an experiment in which its internal research model Claude conducted a campaign to create proteins that bind scorpion neurotoxins with almost no human involvement. The model generated 318,484 variants, of which 211 constructs were synthesized, and binding was confirmed via SPR in two independent laboratories for 17 proteins; the best of them work with nanomolar affinity. This does not amount to an antidote—neutralization of the toxins was not tested—but Anthropic made the report and all laboratory data open.

image
image

What happened

The experiment is structured as a full scientific campaign managed by the model. After receiving the initial task, Claude itself found the structures of five neurotoxins (Cn2, Ts1, AaH I, AaH II, LqhIII) in open PDB and AlphaFold DB databases, selected binding sites, set its own quality thresholds, and combined 12 specialized design tools. It generated 318,484 variants, performed co-folding for 553,052 pairs, and selected 180 constructs for synthesis; humans added another 31 to the list, bringing the total to 211. Then physics came into play: laboratories Adaptyv Bio and Twist Bioscience independently tested all 211 proteins via SPR using a strict criterion—KD ≤ 1 µM, confirmation of the result in both laboratories, and the absence of off-target signal. Binding was confirmed for 17 constructs: 7 against Cn2, 4 against AaH II, 4 against AaH I, and 2 against LqhIII; no candidate passed against Ts1. The two best binders, Cn2-B1 and Cn2-B2, bind the Cn2 toxin with affinity in the single-digit nanomoles. A separate line—91 multi-toxin candidates targeted at multiple targets simultaneously—all failed, with none binding more than one toxin. Claude prepared the final report on the campaign itself, and human involvement in the generation and selection stages was minimal.

Context

Understanding the result helps in distinguishing two goals: binding the toxin and neutralizing it. A binding protein—a binder—is a molecule that "sticks" to the target, and the measure of contact is the dissociation constant KD. For practical application, this is not enough: the candidate must neutralize the toxin, and such a stage was not performed in the campaign. The verification of binding here is stricter than usual in corporate AI reports: instead of the model's self-assessment, the constructs were run through the physical method of SPR (surface plasmon resonance) in two independent laboratories with pre-defined success criteria—a rare standard of proof for this industry. The second layer of context is the change in the role of the language model. Claude here is not an assistant answering scientists' questions, but an executor: it built the plan itself, maintained quality thresholds itself, assembled a working pipeline from specialized tools itself, and wrote the final report. It is also important that this is a one-off research run, not a deployed service: outside the published experiment, Anthropic has no API, pricing, or operational metrics. This recipe—an autonomous agent plus external laboratory verification—is a methodological precedent that other teams can now replicate and verify.

Why this matters for the industry

For the AI industry, the main signal is not the molecules, but the economics of the campaign. Generation and primary selection of candidates fit into approximately 64 hours of autonomous agent work and about 10,000 accelerator hours, and human involvement was reduced to one initial request and four clarifications. If such campaigns become repeatable, the bottleneck will shift from the computable phase, which is scaled by accelerators, to the physical world—synthesis, SPR validation, and functional tests; that is where the next value chain will be concentrated, up to design-as-a-service models at biotech vendors. Developers care about the product pattern itself: "brief → autonomous work → external validation → report": it can be transferred to other domains along with review-gate practices and software acceptance of laboratory orders. Biotech teams get a grounded benchmark: their design pipelines can be compared on the same five toxins with published SPR results without reproducing the campaign from scratch. If the scenario develops, multi-day cycles of "target → candidates → synthesis → measurement" may become a working routine of computational biology and reduce the cost of early stages of therapeutic protein development.

Why this matters for users

For the reader, this is a specific, verifiable example of what agentic models are already doing real scientific experiments, not just writing code and texts. One should not expect an imminent drug from this: the campaign is a one-off research run without API, pricing, or latency data, and the binders are not confirmed as neutralizers, meaning they are not an antidote. The practical value today lies in the openness of the data. The report and technical PDF from Anthropic are published publicly, and the Anthropic/scorpion-toxin-binder-design dataset on Hugging Face is available under the CC-BY-4.0 license and contains all 211 synthesized constructs along with SPR results, including the 17 confirmed binders. With this data, one can verify loud claims about AI protein design, build leaderboards and tools for model evaluation, study the candidate selection methodology, and the approach to agentic scientific campaigns. For a non-specialist, the conclusion is modest: AI agents are entering laboratories, but for now in the role of an accelerator for early research stages, not a creator of finished drugs.

What is still unknown / limitations

Key unknowns are noted in the report itself. Binding of the toxin is not yet neutralization: the neutralizing effect was not tested on cell models or animals, and Anthropic does not call the result an antidote. The formula "about 8% success is comparable to typical metrics of classical protein design" is an assessment by the report's authors, not a correct benchmark: classical campaigns have different targets, hit criteria, and number of attempts, and direct comparison without aligning conditions does not stand up to criticism. External laboratories verified the binding of finished constructs, but the autonomous agentic pipeline itself has not yet been independently reproduced by anyone—this is a one-off demonstration of one model on five specific toxins, and the transferability of the approach to other targets has not been shown. Finally, likely development scenarios—a wave of replications, leaderboards based on the open dataset, design-as-a-service—remain interpretations: there is no confirming data for them yet.

Sources

Author

Look at AI, editorial team