IA · 24 September 2026 · 5 min read
Anthropic steps into biology: Claude uncovers a CRISPR-like enzyme system in just 21 hours
In brief: Anthropic has announced that its Bay Area wet biology lab identified a novel enzyme system with properties reminiscent of the CRISPR gene-editing mechanism. Driven almost entirely by a fleet of roughly 950 Claude agents running for 21 hours, the model pinpointed unusual genomic sequences within bacteriophages that human scientists then confirmed in the laboratory. The milestone illustrates how frontier AI developers are shifting from text generation toward empirical scientific discovery.
by Team Mocchi's
Moving from server racks to wet lab benches
Over the past year, the artificial intelligence sector has been searching for ways to transcend text drafting and software engineering to demonstrate measurable impact in the physical sciences. Following the recent confirmation that it operates an in-house wet biology lab in the San Francisco Bay Area, Anthropic has revealed its first major scientific finding. As reported by TechCrunch, the laboratory identified a previously unknown enzyme system with biological characteristics reminiscent of the renowned gene-editing mechanism CRISPR.
Originally identified as a defense mechanism used by bacteria to fight off viral invaders, CRISPR has become a cornerstone of modern biotechnology by allowing researchers to cut and edit genomic sequences with surgical precision. The enzyme mechanism unveiled by Anthropic was identified within the DNA of bacteriophages—viruses that infect bacteria—and early observations indicate that it can perform operations akin to cutting, copying, and pasting genetic code. While independent peer review will ultimately determine how novel or transformative the system actually is, the announcement highlights a dramatic shift in how scientific hypothesis generation can be automated.
Nine hundred and fifty agents across twenty-one hours
What sets this finding apart from traditional computational biology is the operational setup Anthropic implemented. According to details shared by The Verge, human involvement was intentionally restricted to formulating the initial research query, providing raw genomic databases, and performing physical bench experiments once candidate targets were surfaced.
The search itself was executed by roughly 950 Claude agents operating concurrently across a massive corpus of DNA records. In just 21 hours, the fleet consumed approximately 210 million tokens while scanning through genomic sequences. One of the agent threads flagged an uncharacterized repeating pattern and routed it for human review. Subsequent physical testing in Anthropic's wet lab confirmed the presence of an active enzyme mechanism within the virus sample, turning a computational signal into a verified biological observation in a matter of days.
Scientific caution meets the enterprise AI race
Despite the excitement surrounding the announcement, the broader scientific community is approaching the findings with measured skepticism. Anthropic CEO Dario Amodei acknowledged on social media that the discovery builds on prior academic work, pointing out that researchers at Stanford University had previously described related molecular machinery. It remains entirely open whether this particular enzyme can be repurposed into practical biotechnological tools or compete with existing gene-editing platforms.
The strategic context surrounding the announcement is impossible to ignore. As capital expenditure for frontier models soars and leading labs prepare for eventual public market debuts, proving concrete utility in hard sciences like drug discovery and therapeutics has become essential. While OpenAI has frequently highlighted advances in formal mathematics, Anthropic is placing its bets on molecular biology to recruit top scientists and forge long-term research alliances with global biotechnology leaders.
Mocchi's take
This development illustrates the transition of agentic artificial intelligence from desktop productivity tools to powerful combinatorial search engines for empirical science. For software teams and enterprise decision-makers, the real takeaway lies in multi-agent orchestration: deploying nearly a thousand task-focused agents to parse vast datasets compressed what used to take years into less than a single day of compute. Equally important, the requirement for physical lab validation serves as a reminder that software cannot replace domain expertise—instead, it acts as an unprecedented multiplier for rigorous real-world testing.