Anthropic said its Claude model autonomously discovered a previously uncharacterized enzyme system in bacteriophages, the first published result from the company's newly-launched wet lab. It took nearly 950 Claude agents working in parallel for 21 hours, burning through 210 million tokens of DNA sequence data, before one agent flagged an unusual repeating pattern for human review. Anthropic is comparing the machinery to the enzyme systems behind Crispr, the gene-editing tool that reshaped biotechnology over the last decade.
Human scientists' involvement was limited to writing the initial prompt and running the follow-up bench work, according to the company. Everything between — the search across a massive DNA sequence database, the pattern recognition, the shortlisting — was done by the agent swarm. Further analysis and lab testing then confirmed the flagged pattern corresponded to a real, previously undescribed enzyme system in phages, the viruses that infect bacteria.
The 210-million-token compute budget is the load-bearing number here. It reframes what agentic search on scientific corpora costs in practice: not a single long-context query but a distributed sweep across nearly a thousand agents, each burning roughly 220,000 tokens on average. That's the scale it took to surface one novel finding from a genomics dataset, and it hints at what Anthropic thinks the unit economics of AI-led discovery look like today.
Key facts
- 01Nearly 950 Claude agents ran for 21 hours and consumed 210 million tokens before flagging the pattern.
- 02The find is a previously uncharacterized enzyme system in bacteriophages, viruses that infect bacteria.
- 03Anthropic says scientist involvement was limited to the initial prompt and follow-up lab work.
- 04It is the first result from Anthropic's newly-launched wet lab.
- 05Anthropic is comparing the enzyme system to the machinery behind Crispr, though practical applications remain unclear.
Anthropic conceded it is still working out what the enzyme system actually does. The Crispr comparison is structural — a repeating pattern of biological machinery associated with phage defense — not functional. Whether the system can be turned into a usable tool, or whether it will remain a curiosity in the phage literature, is not yet established.
The company said it chose to publish anyway. In its own words, Anthropic felt it was important to share such findings early, both to demonstrate Claude's capabilities and to give the broader community insight into what we're working on. That is a departure from standard biology publishing norms, where discoveries typically wait for functional characterization and peer review before going public.
“important to share such findings early, both to demonstrate Claude's capabilities and to give the broader community insight into what we're working on.”— Anthropic, company statement
The timing is not incidental. Anthropic is preparing to go public and is actively recruiting scientists to its lab, with drug discovery among the stated expansion targets. A headline-grade biology result — even a preliminary one — is exactly the kind of proof point that shifts an AI lab's story from chatbot vendor to research infrastructure. The wet lab itself, launched to pair Claude with real experimental capacity, is central to that repositioning.
The broader pattern across frontier labs is the same. OpenAI has been pushing Claude and GPT-class models against olympiad-level mathematics, sometimes cutting corners on the academic conventions of the field to publish results faster. Google DeepMind's AlphaFold franchise remains the reference point for what AI in the life sciences can look like at maturity. Anthropic's phage-enzyme announcement is a bid to plant a flag in that same territory, on its own terms.
The caveats are real. A structural resemblance to Crispr is not a functional one, and the history of biology is thick with proteins that looked promising in silico and did nothing useful at the bench. Anthropic itself is careful not to claim therapeutic or industrial relevance. Whether independent labs can reproduce the finding, and whether the enzyme system does anything Crispr-adjacent, is the test the announcement will ultimately be judged on.
For the AI market, the more durable signal is the compute recipe. Nearly 1,000 agents, 21 hours, 210 million tokens, one novel biological pattern surfaced — that's a template competitors can copy, and it puts a rough price tag on autonomous scientific search. If the finding holds up and more like it follow, Anthropic will have shown that agent swarms plus a wet lab is a workable pipeline for turning frontier model spend into publishable science, and that changes the calculus for every lab weighing whether to build its own.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
