Anthropic says its new wet biology laboratory has identified a previously unknown enzyme system with properties resembling CRISPR. The company attributes much of the discovery process to Claude, but the significance and novelty of the result still require validation by researchers outside Anthropic.
The system was found in bacteriophage DNA. Anthropic’s researchers say it can perform operations comparable to cutting, copying and pasting DNA, language that suggests a potentially useful biological mechanism without establishing a finished gene-editing technology.
CRISPR began as a natural bacterial defense against viruses and became a widely used research tool for editing genes. Anthropic’s comparison helps describe the claimed behavior of its enzyme system, but resemblance to CRISPR does not by itself establish equivalent precision, safety or practical value.
Claude Searched With Hundreds of Agents
Anthropic says Claude reached the result after 21 hours of concentrated work. About 950 agents searched data and consumed 210 million tokens during the process. Chief executive Dario Amodei described the discovery as mostly, though not entirely, made by Claude.
The speed is striking because the Bay Area laboratory was established only this spring. Yet the company has not claimed that the model independently completed the whole scientific process. Human scientists performed the physical experiments, and the model’s role centered on analysis and directing the search through existing information.
The token count and number of agents indicate the computational scale of the search, not the quality of the biological conclusion. Anthropic has not provided a comparison showing how long the same task would have taken a conventional research team or how often similar agent searches fail to produce useful candidates.
Anthropic also acknowledges that the discovery builds on prior work. Amodei said a Stanford team had already found a system similar in some respects. That history and the need for external replication limit how confidently the result can be described as entirely new or transformative.
Physical Experiments Remain Under Human Control
The laboratory handles only lower biosafety levels, BSL-1 and BSL-2, and Anthropic says it does not work with pathogens capable of infecting humans. All current physical laboratory work is performed by human researchers rather than autonomous machines.
That boundary is important because Amodei has warned that powerful AI could be misused for bioterrorism. At the same time, he predicts that AI could help cure most diseases within five to ten years. Both statements are his projections, not outcomes established by the reported enzyme work.
The timing makes that tension especially visible. Anthropic disclosed the laboratory soon after executives and employees publicly emphasized severe model risks and the need for stronger safety testing. The company has concluded that the possible benefits justify proceeding under its current controls, but TechCrunch does not independently assess those safeguards.
Amodei has not ruled out a future in which Claude safely operates laboratory equipment with appropriate safeguards. Anthropic says that is not happening today. The present setup therefore combines large-scale automated reasoning with human control over instruments, materials and experiments.
The Claim Now Moves to Scientific Validation
Researchers outside Anthropic will have to determine whether the enzyme system is as novel and useful as the company suggests. The TechCrunch report does not include peer review, independent replication or a demonstrated medical application. It presents an announcement with clearly stated limits.
Anthropic is also not alone in applying AI to biology. The source points to Stanford work combining language models and CRISPR, University of California San Francisco researchers using AI to design enzymes, and Google’s earlier AlphaFold system. Those examples show an active field rather than a discovery pathway unique to one company.
No clinical claim follows from the announcement. The reported system exists at the stage of a laboratory finding, and any medical use would require evidence not contained in the source article. External scientists may ultimately confirm, narrow or reject Anthropic’s interpretation.
The central development is the scale and organization of Claude’s contribution: hundreds of agents conducting an intensive data search before humans tested the resulting idea. Whether that workflow becomes a repeatable research method will depend on independent confirmation and on whether future results can survive the same scrutiny applied to conventional laboratory science.
