Claude found a CRISPR-like pattern; humans still had to turn it into biological evidence

Anthropic says Claude identified a previously unknown viral enzyme system resembling the machinery behind CRISPR. The search used about 950 agents, 210mn tokens and 21 hours of computation to examine bacteriophage DNA and propose candidates. Human researchers then performed the physical experiments in Anthropic’s new Bay Area laboratory.

That division of labour is the important result. The agents could search literature and sequence data, form hypotheses and coordinate a large computational investigation. They did not autonomously operate the laboratory, and the biological claim exists in a preprint rather than a peer-reviewed paper. The system’s repeated DNA pattern resembles CRISPR arrays, but resemblance and initial activity do not make it a safe, programmable editor for human cells. Researchers still need to establish mechanism, targeting accuracy, efficiency, off-target effects and the range of organisms in which it works.

The experiment nevertheless advances the evidence beyond a model writing plausible research plans. A wet lab creates an external test that language cannot bluff: a proposed molecule either produces the measured effect or it does not. The 950-agent scale also exposes the economics. Two hundred and ten million tokens plus laboratory staff and equipment may be worthwhile for a genuinely novel lead, but the relevant comparison is with expert researchers and existing bioinformatics pipelines—not with a single chat session.

Anthropic’s wording should therefore be read literally. Claude helped discover a possible new gene-editing system. If independent groups reproduce and develop it, the work may become a valuable tool. Today, the stronger conclusion is methodological: agents can compress a broad search into a laboratory-testable hypothesis, while humans and experiments remain the source of biological validation.