Claude discovers an unknown enzymatic system in 21 hours: Anthropic announces the first hit from its biology lab

Claude discovers an unknown enzymatic system in 21 hours: first hit for Anthropic's biology lab

🔎 An enzyme unknown to science, spotted by a fleet of agents

On September 24, 2026, Anthropic announced a result that goes beyond the usual run of AI press releases: its Claude agents identified an enzymatic system never before described by science. The whole thing took 21 hours, with 950 agents sifting through 200,000 enzyme sequences — roughly 210 million tokens in total. The system, dubbed ART (array-associated reverse transcriptases), features a structural architecture closely resembling CRISPR arrays. According to Anthropic, it may even be potentially programmable.

This is no stylistic exercise: it's the first concrete result from the molecular biology lab the company quietly set up. And it comes with heavyweight validation — Feng Zhang, a CRISPR pioneer, finds the work "genuinely intriguing." Dario Amodei, Anthropic's CEO, even cited the discovery before the UN Security Council.

Between the marketing hype ("the next CRISPR") and the scientific reality (unknown function, no demonstrated applications), there's a gap. Our job: to measure it.

Key takeaways

• 950 Claude agents scanned 200,000 enzyme sequences in 21 hours (210 million tokens), narrowing 3,500 candidates down to 20 promising systems.
• One of them, the ART system (array-associated reverse transcriptases), is structurally close to CRISPR arrays and "potentially programmable," according to Anthropic.
• Unknown function: no therapeutic application has been demonstrated to date. Caution is warranted.
• Feng Zhang (Broad Institute) calls the work "genuinely intriguing" — a measured compliment, not a verdict.
• Dario Amodei cited the discovery before the UN Security Council: the subject has become geopolitical.
• This is Anthropic's biology lab's first hit, built precisely to close the loop from hypothesis → experimental validation.

Recommended Tools

To understand this type of announcement, verify it, or launch your own exploration pipeline, here is the minimal toolbox:

Anthropic's discovery rests on this foundation: public data, mature comparison tools, and above all an agent layer capable of sustaining biological reasoning at industrial scale.

What exactly did Claude discover? An ART system, cousin of CRISPR arrays

Direct answer: a system pairing a reverse transcriptase with an array of repeated sequences — an architecture that immediately brings CRISPR to mind, without anyone yet knowing what it's for.

A useful refresher. A CRISPR array is bacterial memory: regular repeats of DNA, separated by fragments of viruses attacked in the past. In the ART system, the same array logic is found, but the associated protein is not a nuclease: it's a reverse transcriptase, the enzyme that copies RNA into DNA. Hence the name: array-associated reverse transcriptases.

The "potentially programmable" qualifier deserves to be weighed word by word. If this system uses the sequences in its array as guides to read or write DNA, one could imagine hijacking it into a tool — targeted editing, biological recording, molecular tracking. But that's a working hypothesis, not a result. Prime editing, one of the most promising editing techniques of the decade, already relies on a reverse transcriptase guided by an RNA guide: enough to understand the sector's sudden appetite for these architectures.

The crucial point, meanwhile, comes down to one sentence: the function of the system is unknown. No one yet knows what it does in the host cell. The 20 systems selected from 3,500 candidates are not all equal, and ART is the star of the shortlist — not a ready-made solution.

From observation to tool: 25 years for CRISPR

History explains why these architectures fuel dreams — and why patience is required. CRISPR repeats were described as early as 1987 in the genome of Escherichia coli. Their immune defense function was only demonstrated in 2007. The programmable toolkit arrived in 2012–2013. Between the curious observation and the tool that changes medicine: a quarter of a century.

Even in an optimistic scenario where AI agents compress the exploration phase, ART's trajectory looks more like the beginning of a quest than an arrival. Bacterial reverse transcriptases — retrons, DGR elements — have already provided biotech building blocks, and some CRISPR systems themselves incorporate RTs. The ground is fertile. Fertile is not cultivated.

950 agents, 21 hours, 210 million tokens: the real story of the run

Direct answer: Claude didn't have an epiphany — it industrialized a sorting task that no one could do by hand.

The rundown, as reported by Al Jazeera and Phys.org: a database of 200,000 enzyme sequences; 950 agents launched in parallel; 210 million tokens consumed in 21 hours; 3,500 candidates identified, then narrowed down to 20 systems after cross-checks and scoring. Then — and this is the part the headlines forget — biologists, a lab, experiments.

What do the agents actually do?

No magic here: cross-referencing the literature, predicting architectures, generating functional hypotheses, and above all self-critique — one agent proposes, another attacks the hypothesis. The funnel is brutal: 3,500 candidates down to 20 survivors, an elimination rate of 99.4%. It's precisely this high rejection rate that makes the result credible: the fleet didn't validate everything, it pruned.

My take: the real story of this announcement isn't the intelligence, it's the logistics. Microbial genomes are a giant junk drawer — millions of genes of unknown function. Sorting through that junk with reasoning, and not just pattern matching, is exactly what LLM agents do well. As for the cost of the run — a few thousand dollars of tokens at public rates depending on the model — it's negligible compared to a lab campaign.

It's no coincidence that this result is coming out now: Anthropic quietly set up a physical biology lab precisely to close the loop between AI-generated hypotheses and their experimental validation. The lab isn't a PR gimmick; it's what makes everything else credible.

Agent fleets: scientific discovery becomes a logistics problem

Direct answer: this discovery is not an isolated event, it's the umpteenth iteration of a pattern that is taking hold across every field.

Three recent examples, one single mechanism:

• Cybersecurity: Claude Mythos discovered over 10,000 critical vulnerabilities in one month. Massive fleet, automatic triage, escalation to humans for remediation.
• Mathematics and physics: OpenAI claims to have solved Navier-Stokes in 88 hours with 10,000 agents. Verifiable formal proof, narrow scope — but the same fleet architecture.
• Tooling: Claude Code Agent View, the dashboard for steering dozens of agents in parallel, shows that fleet orchestration is becoming a product in its own right.

Under the hood, there's also the question of compute: the Anthropic-SpaceX contract for Colossus 1, with its 220,000 GPUs and 300 MW, is a reminder that these agent runs are backed by massive energy industrialization.

The pattern is always the same: parallelize the exploration, filter, have humans validate. What's really changing is the bottleneck. It is shifting from "finding hypotheses" to "proving hypotheses." Fleets produce candidates on an assembly line; labs, instruments, and experts become the scarce resource. Anthropic understood this before everyone else by building its own.

"Genuinely intriguing": Zhang applauds, the community awaits the data

Direct answer: Feng Zhang's quote is a measured compliment, and it should be taken exactly as such.

Feng Zhang is no random commentator. A researcher at the Broad Institute and MIT, he is one of the pioneers of CRISPR-Cas9 editing in mammals. When he calls the work "genuinely intriguing" — sincerely intriguing — he is praising the quality of the analysis, not heralding a therapeutic revolution. Intriguing, period. In a field where superlatives fly fast and loose, this restrained vocabulary is a signal of seriousness on both sides.

The sober reading hinges on one distinction: structural similarity does not equal function. The history of bacterial defense systems is full of elegant architectures described one year, then forgotten for lack of a demonstrated mechanism. The proof will come in three steps: demonstrating what the system does in its natural context, demonstrating that it can be reprogrammed, and only then talking about applications.

What would real validation look like?

Concretely: in vitro reconstitution of the system, characterization of its molecular product, reverse genetics to confirm its role in the host bacterium, then a high-resolution structure. Months of work, dedicated teams, expensive instruments. This is exactly the phase for which Anthropic has a lab — and that is where the announcement will be judged.

In the meantime, the hype machine is already running. Phys.org notes Anthropic's aggressive communications push around its "discovery in biology," and early analyses are already brandishing the "next CRISPR" without the slightest evidence of a therapeutic application. Let's not forget a well-understood self-interest: Anthropic sells Claude. Every scientific discovery is also a commercial demonstration. That doesn't invalidate the result — it should simply calibrate your enthusiasm.

Amodei at the UN: AI-assisted biology changes scale

Direct answer: when the discovery of a bacterial enzyme gets cited at the UN Security Council, it means the debate has shifted ground.

Dario Amodei cited this result before the Security Council, according to coverage of the announcement relayed by AIweekly. Two readings, and they're not mutually exclusive. The showcase: AI as an engine of beneficial discovery, capable of expanding our knowledge of living systems. The framing: AI-assisted biology as a collective security question, where the same agents that identify unknown systems could, in the wrong hands, serve other ends.

The industry's response to the risk is itself agentic: OpenAI launched Rosalind, an agent dedicated to biodefense to prepare the response to biological threats. Discovering and protecting with the same tools: that's the narrative the two labs are building in parallel, each with its flagship model.

My take: this UN moment says more about the Anthropic-OpenAI race than about biology. Each player must now prove two things at once — that its AI discovers, and that it secures. Yet the benefit/risk ratio of a system with unknown function is, by definition, unknown. This is exactly the kind of object that Responsible Scaling Policy-style frameworks are supposed to govern before any scale-up, and it will take real rigor to hold that line when the headlines start talking about "the next CRISPR."

Claude facing the competition: who's leading the race to agent-driven science?

Direct answer: on paper, in agentic benchmarks, OpenAI dominates; on vertical integration with a real lab, Anthropic has just taken a symbolic lead.

These scores, drawn from the latest public agentic rankings, don't capture everything. Anthropic's differentiation in this matter isn't a number: it's the complete loop — models, agents, a physical lab, a network of experts capable of making the final call. OpenAI is responding with Rosalind on the biodefense side; Google is betting on long-form reasoning with Deep Think. Three strategies, one shared observation: value is shifting from the model to the system that surrounds it.

To choose a model for your own agentic workflows, our guides remain useful: our Claude vs ChatGPT comparison for the head-to-head, and our selection of the best LLMs for coding — code remains the training ground for agents, and it's there that their reliability and real autonomy are judged.

What This Changes — and What It Doesn't — for Research

Direct answer: AI cuts the cost of exploration by a factor of hundreds; it changes almost nothing about the cost of proof.

On the exploration side, the break is stark: sorting through 200,000 sequences used to mean several months of methodical postdoc work. The run did it in 21 hours, with coverage and consistency no human team can match at that volume. On the proof side, nothing has moved: characterizing an unknown function requires cultures, biochemistry, genetics, replicates — months, sometimes years. Anthropic's discovery doesn't shorten that phase; it makes it more targeted, and therefore less wasteful.

At bottom, this is good news for biologists rather than a threat. The agents produce candidates; humans decide what deserves six months at the bench. Biology isn't being automated — it's being fed. The shortage taking shape is not one of ideas, but of expert hands and of machines to run the tests.

❌ Common Mistakes

Mistake 1: Believing that "Claude made a discovery all on its own"

What's wrong: the headlines compress a human-machine pipeline into a myth of autonomous AI. The agents did the sorting; biologists framed the targets, designed the validation checks, and interpreted the results. The fix: read "AI-assisted" whenever a press release says "AI discovery". The nuance is the whole point.

Mistake 2: Talking about "the next CRISPR"

What's wrong: no therapeutic application has been demonstrated, the function of the ART system is unknown, and structural similarity alone does not make an editing platform. The fix: wait for the functional data — a demonstrated mechanism, then demonstrated reprogramming — before investing in the word "revolution". CRISPR itself took 25 years from observation to tool.

Mistake 3: Comparing runs on raw numbers

What's wrong: "950 agents, 210 million tokens" is not a quality metric. Without a shared protocol, these numbers mostly serve as comparative marketing between labs. The fix: watch the rate of experimentally validated candidates in the coming months. That's the only KPI that matters, and it arrives with a delay — as always in biology.

❓ Frequently Asked Questions

What is an ART system?

ART stands for array-associated reverse transcriptases: reverse transcriptases associated with an array of repeated sequences in bacterial genomes. The system identified by the Claude agents is structurally similar to CRISPR arrays and considered potentially programmable by Anthropic. Its actual biological function, however, remains unknown at this stage, which calls for caution.

Has the discovery been experimentally validated?

The computational screening is complete; functional characterization is only just beginning. Feng Zhang describes the work as "genuinely intriguing," which validates the value of the analysis, not its application. No therapeutic application has been demonstrated, and definitive results will require months of laboratory work, with the usual uncertainties of research.

How much does a run of 950 agents cost?

210 million tokens in 21 hours represents, at public API rates, on the order of a few thousand dollars depending on the model and the input/output mix. That is negligible compared to the cost of an experimental campaign. The real budget for this type of program lies in validation, not in computational triage.

Does Claude replace biologists?

No. The agents compress exploration and propose candidates; humans design the experiments, interpret the results, and bear the scientific responsibility. The physical laboratory Anthropic built exists precisely because a model, on its own, proves nothing: it produces hypotheses, and only hypotheses.

Where can I verify this information?

The reference coverage dates from September 24, 2026: Al Jazeera, Phys.org, and AIweekly. Technical details can be found in the team's publication. To judge the real significance of the discovery, wait for the functional data: that's where the rest of the story will unfold.

✅ Conclusion

In 21 hours, 950 Claude agents transformed a molecular junk drawer into a shortlist of 20 systems, including one serious candidate: the discovery is real, the revolution will wait for proof. To see how Claude stacks up against GPT-5.5 and Gemini 3 Pro Deep Think, head over to our Claude vs ChatGPT comparison.