Build Reports disagree 5 publishers 2 min read Published Sep 27, 2026, 3:17 AM PDT

Anthropic's 950 Claude agents spent 210 million tokens to hand biologists 20 enzyme leads

Anthropic ran about 950 Claude agents for 21 hours on 210 million tokens to cut over 200,000 reverse transcriptases to 20 lab candidates. Human scientists kept the bench work, so lab capacity sets how fast those leads get tested.

The Engineer · Build desk

Send a correction

Illustration accompanying Anthropic's 950 Claude agents spent 210 million tokens to hand biologists 20 enzyme leads — Generated illustration

What happened

Between the raw haul and the shortlist, the agents flagged about 3,500 candidate systems, according to Anthropic's account.
One agent noticed repeating DNA beside an unusual reverse transcriptase gene, a pattern the researchers had not asked it to look for.
Anthropic named the system ART, and its early lab experiments found the repeat array produces short RNAs whose function is still unknown.
Earlier studies had already described the reverse transcriptase itself, so the claim rests on the repeats and a partner gene forming a larger system.
Lucas Harrington, a Mammoth Biosciences co-founder, wrote on X that the method is decades old and that similar systems have been known since 2008.

Why it matters

What it costs cost Each of the 20 finalists took about 10.5 million tokens of agent work before any lab time, and a team copying the design pays roughly that per lead.
What it limits constraint Because humans run every experiment, adding agents or tokens raises the number of leads without raising how many of them the lab can test.
Who has to decide decision Teams planning an agent search now have a worked split to copy or reject: per Dario Amodei, humans picked the search area and ran the wet lab, and agents handled the triage.
What doesn't add up contradiction Anthropic presents ART as a discovery, while Harrington argues the hard part is showing what a system does, and ART's function has not been shown yet.

Next story · Science · Sep 24

950 Claude agents narrowed 200,000 enzyme genes to 20 leads in 21 hours

Confirmed2 publishers

Spread across every enzyme gathered, 210 million tokens comes to at most about 1,050 tokens per reverse transcriptase ^[13]. Spread across the flagged candidate systems, it is about 60,000 tokens each ^[14]. Per agent, the run averaged about 221,000 tokens ^[16]. The fleet used about 10 million tokens an hour ^[17].

Anthropic reported only totals for time and tokens ^[1]. The accounts give no per-stage split, no model version and no price.

Those averages describe Anthropic's workload. For them to transfer, the target has to be a sequence database big enough that a pass over every record is worth paying for, and a lab has to be ready to take the finalists. The figures are also Anthropic's description of its own system, in a preprint that has not completed peer review ^[10].

The design choice I would defend is the handoff. The agents wrote a human-readable report on each finalist ^[4]. Scientists read those reports, and the 3,500 flags stayed with the agents ^[3]. The physical experiments happen in the Bay Area lab Anthropic built after forming its life sciences group in spring 2026 ^[9].

The 2024 SAMPLE system closed the loop differently. It sent its protein designs to automated lab equipment and used the results to pick the next test, all aimed at better heat tolerance ^[12]. SAMPLE was aimed at a fixed target, while Anthropic's agents worked from an open brief ^[5].

A brief to look for interesting enzymes would fail any spec review I have sat in ^[5], and it can also surface a pattern nobody named in advance. Before escalating the ART lead, the agent checked the repeat spacing, compared the system with known biology and searched published research ^[6].

CRISPR researcher Feng Zhang said the finding is "intriguing and merits further investigation" after reviewing the preprint, The Neuron reported ^[11]. On this evidence, the new part is that agents ran the triage stage of a method genome miners already use ^[19]. Finding out what ART's short RNAs do is bench work ^[8], and humans do the bench work ^[9].

What to watch

Peer review of the ART preprint, and whether outside groups can reproduce the agent search with their own models and data.
Lab results showing what ART's short RNAs do, the gap Lucas Harrington says Anthropic has not closed.
A per-stage token breakdown or compute bill from Anthropic, to show whether the first pass or the triage used most of the 210 million tokens.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence46
Adoption: Not scored
Hype gap+48
Incentives80
Confidence64

Perspective Coverage

5 publishers

Builder: Builder 45%
Operator: Operator 36%
Investor: Investor 19%

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
According to Anthropic, about 950 Claude agents searched for roughly 21 hours and used 210 million tokens.
Reported Supported Source: Anthropic, as reported by The NeuronView cited source
[2]
The agents collected over 200,000 reverse transcriptases, enzymes that copy RNA into DNA.
Reported Supported Source: Anthropic, as reported by The NeuronView cited source
[3]
From the reverse transcriptases, the agents identified about 3,500 candidate systems and narrowed the list to roughly 20 for closer review.
Reported Supported Source: Anthropic, as reported by The NeuronView cited source
[4]
The agents produced human-readable reports on the 20 most compelling candidates for scientists to review.
Reported Supported Source: Anthropic, as reported in The Neuron newsletterView cited source
[5]
One agent noticed repeating DNA next to an unusual reverse transcriptase gene. Researchers asked Claude to find interesting enzyme systems but did not specifically ask it to search for this type of repeat pattern.
Reported Supported Source: The NeuronView cited source
[6]
The agent reviewed the spacing of the repeats, compared the system with known biology, and searched published research before sending the finding to human scientists.
Reported Supported Source: The NeuronView cited source
[7]
The reverse transcriptase itself was already known from earlier studies, as Reuters reported; Anthropic's discovery claim centers on the repeat array and nearby protein appearing to form a larger biological system.
Reported Supported Source: The Neuron, citing ReutersView cited source
[8]
Anthropic named the system array-associated reverse transcriptases (ART). Its early experiments found ART's repeat array produces short RNAs; researchers still need to determine their function.
Reported Supported Source: Anthropic, as reported by The NeuronView cited source
[9]
Anthropic formed its life sciences research group in spring 2026 and built a molecular biology lab in the Bay Area. Human scientists perform the physical experiments while Claude searches genomic data and proposes candidates for review.
Reported Supported Source: The NeuronView cited source
[10]
The ART research remains a preprint and has not completed peer review; Anthropic is describing the performance of its own AI system.
Reported Supported Source: The NeuronView cited source
[11]
"intriguing and merits further investigation" (Feng Zhang, describing the finding after reviewing the preprint)
Reported Supported Source: Feng Zhang, as quoted by The NeuronView cited source
[12]
A 2024 protein-engineering study introduced SAMPLE, which proposed protein designs, sent them to automated lab equipment and used the results to choose what to test next, working toward enzymes with better heat tolerance.
Reported Supported Source: The NeuronView cited source
[13]
The run used at most about 1,050 tokens per reverse transcriptase gathered.
Derived Supported
[14]
The run used about 60,000 tokens per flagged candidate system.
Derived Supported
[15]
The run used about 10.5 million tokens per finalist candidate.
Derived Supported
[16]
The run averaged about 221,000 tokens per agent.
Derived Supported
[17]
The fleet averaged about 10 million tokens per hour.
Derived Supported
[18]
Anthropic CEO Dario Amodei said Claude mostly led the discovery work while humans chose the search area and ran the wet-lab checks it proposed.
Reported Contested Source: Dario Amodei, as reported in The Neuron newsletterView cited source
[19]
Lucas Harrington, co-founder of Mammoth Biosciences, which uses genome mining to hunt for CRISPR systems, wrote on X that the method has been around for decades and that similar systems have been known since 2008.
Reported Contested Source: Lucas Harrington, as reported by The DecoderView cited source
[20]
Harrington wrote that the hard part is figuring out what a system actually does, that Anthropic hasn't shown that, and that presenting early results as a major discovery isn't helpful.
Reported Contested Source: Lucas Harrington, as reported by The DecoderView cited source

Sources

5 independent publishers whose own reporting we read for this story.

dev.to
1 article · September 24, 2026
Claude Discovers a CRISPR-Like Enzyme System Called ART
economictimes.indiatimes.com
1 article · September 23, 2026
Anthropic: Anthropic touts AI-led biology discovery - The Economic Times
lesswrong.com
1 article · September 24, 2026
Anthropic shares an exciting result in enzyme discovery - and an exercise in public's perception of science and AI
the-decoder.com
2 articles · September 24, 2026
Anthropic says Claude discovered a new enzyme system, but CRISPR researchers call it routine genome mining
theneuron.ai
3 articles · September 25, 2026
😺 What 950 Claude agents found

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder45%

The builder interest is the method, and The Neuron and dev.to carry it in the most detail. Anthropic set roughly 950 agents on a genomic database, gathered more than 200,000 reverse transcriptases, and funneled them to 3,500 candidates and then 20 finalists before one agent flagged a repeat pattern nobody had asked it to look for. The Neuron distills this into a reusable pattern: split a question into parallel scouts with a fixed output shape, then collapse the results with one reviewer. Dev.to is careful that the 3,500-to-20 step, not the raw agent count, is where the model did something closer to expert triage than search.

Operator36%

For anyone running agents in production, the transferable point is where the bottleneck lands. Anthropic says human scientists carried out every physical experiment while Claude searched and proposed; The Neuron and dev.to both stress that division. The Neuron draws the consequence plainly: if agents generate credible leads faster than a lab can test them, wet-lab capacity becomes the constraint and teams need a way to rank machine-generated ideas. The run's figures show just how cheap the search step now is relative to the bench step it feeds.

Investor19%

The Economic Times covers this as an Anthropic milestone, complete with Amodei citing the finding to the UN Security Council and predicting progress on diseases that have plagued humanity for millennia. Weigh that against what the reporting establishes: the result is a self-described preprint that has not been peer reviewed, the underlying enzyme was already known, and the system's function is unresolved. The Decoder gives space to Mammoth Biosciences co-founder Lucas Harrington, who calls the technique routine genome mining and the 'major discovery' framing unhelpful.