What Is Anthropic's Biolab? Claude's ART Enzyme Discovery
What Anthropic announced on September 24, 2026
Anthropic introduced a life sciences research group and laboratory and, in the same post, disclosed its first result: Claude autonomously discovered a novel enzyme system in bacteriophages that the team has named array-associated reverse transcriptases (ART). The company released a technical preprint alongside the announcement.
The group was formed in spring 2026 to test whether general AI models can systematize the kind of discovery that has historically started with a scientist noticing something odd in nature: restriction enzymes in bacterial immune systems, Taq polymerase in a Yellowstone hot spring, CRISPR as an unusual repeat sequence. Anthropic’s premise is that the acceleration comes from agents collaborating with humans at every step, which required building its own lab and a single team spanning Claude training, computational biology and bench experiments.
What ART is
The system is built on a reverse transcriptase (RT), an enzyme that copies RNA into DNA, found in a jumbo phage (a large virus that infects bacteria). The RT itself was already in the literature. What Claude noticed, and what no prior study had recorded, were two features sitting next to it:
- An array of evenly spaced non-coding DNA repeats, a pattern reminiscent of the CRISPR arrays that store RNA guides in gene-editing systems.
- An accessory protein of unknown function.
Anthropic’s careful claim is that this combination of characteristics has only ever been found together in a handful of other systems, all of which are programmable and perform operations like cutting, copying and pasting DNA. Beyond CRISPR itself, several such systems are in development as next-generation editing tools. The function of ART is not yet known; characterization is ongoing, and Anthropic says it is publishing early to show Claude’s capabilities and to let the community see what it is working on.
How the discovery happened
The workflow is the part most relevant to anyone building agent systems:
| Step | What happened |
|---|---|
| Prompt | Scientists asked Claude to search a massive DNA database for interesting new examples of RTs. That was the extent of human direction before the lab stage. |
| Scale | Roughly 950 agents, 21 hours, 210 million tokens. |
| Method | Agents surveyed the RT protein family, read the relevant literature, reproduced established results from public data to validate their methods, then hunted for family members or genomic neighbors that fit no described system. |
| Filtering | Each candidate got a short human-readable report proposing a function and citing evidence; a follow-up critical-evaluation pass eliminated most candidates. A survey may end with one candidate or none. |
| Hit | One agent flagged a repeating DNA pattern next to the gene for an odd-looking RT. |
| Validation | Human scientists tested the candidate in the wet lab and confirmed a previously uncharacterized system. |
Third-party reporting puts the search at about 1.9 billion protein clusters and roughly 3,500 candidates narrowed to around 20 systems for follow-up; Anthropic’s own post gives the 950-agent, 21-hour, 210M-token figures.
About the lab
The lab is in the San Francisco Bay Area and “looks like a typical molecular biology lab.” It works only at BSL-1 and BSL-2 and does not handle pathogens that infect humans. All bench work is performed by human scientists; Anthropic has experimented with AI-driven lab automation through its Model Hardware Standard research preview but says that approach suits ad hoc molecular biology workflows poorly. The team’s prior publications cover CRISPR evolution and regulation, new enzymes for cell and gene therapies, and tools for identifying pathogenic human variants.
The group is part of a larger life sciences organization at Anthropic that also includes drug discovery and the training of Claude in biology and chemistry.
Why the timing matters
The disclosure landed two days after Claude Opus 5.5 shipped and one day after Dario Amodei told the UN Security Council that AI could pose a risk to humanity, with Anthropic’s IPO scheduled for November 2026. It is also Anthropic’s most concrete evidence yet for the biology thesis in Amodei’s “Machines of Loving Grace” essay, and it arrives as the company’s R&D Automation Index tries to quantify how much research Claude does on its own.
Feng Zhang, one of the pioneers of CRISPR genome editing, reviewed the preprint and called it “an exciting example of how AI agents can contribute to biological discovery,” adding that the RNA-repeat arrays “merit further investigation.”
What to take from it
- For biologists: a candidate programmable system with CRISPR-like architecture, function unknown, preprint available. Worth watching, not yet worth citing as a tool.
- For agent builders: the loop that worked was survey → reproduce known results → hunt for anomalies → write a report per candidate → critically re-evaluate → hand off to humans. The reproduction step and the self-critique step are what kept 950 agents from flooding the lab with false positives.
- For everyone else: this is the first named discovery from a frontier lab’s in-house wet lab, and the first where the AI’s contribution was the noticing rather than the analysis.
Last verified: September 25, 2026.