Anthropic Moves Into Biology: What Is Announced vs Shown

Between June and September 2026, Anthropic, the company behind the Claude models, made four moves into the life sciences. It launched a research workbench called Claude Science on June 30. It signed a collaboration with Novo Nordisk, announced on September 16. A day later it opened a gated access program for biology work. And on September 23 it reported that its agents, working with an internal wet lab, had found an enzyme system nobody had described. Taken together, it reads like an AI company turning itself into a biotech supplier. Read one by one, the four announcements carry very different amounts of evidence.

The access program: rules, not results

The Life Sciences Verification Program, announced on September 17, changes what researchers are allowed to ask. Anthropic's models normally block many biology requests in real time, because the same knowledge that helps a vaccine lab can help someone build a weapon. Verified teams get access to Mythos, Opus and Sonnet with more permissive biology safeguards. The beta is open to teams and institutions only, not to individuals, and Anthropic says dozens of organizations were already admitted in early access.

There are two tiers. A Standard Use grant is renewed every year and covers Mythos 5.1, Opus 5 and Sonnet 5. A High risk Use grant is approved project by project, renewed every six months, and lifts all life sciences blocks. It is available on Opus 5 and Sonnet 5; for Mythos, Anthropic limits it to a small number of entities and says it is working with the US government. Cyber classifiers stay on in every tier.

The trade is visible in the fine print. Real time blocking is replaced by offline monitoring, with conversations kept for 30 days. Anthropic says that data is walled off, not used for training, and not accessible to its own life sciences research teams. Vetting checks research credentials, security and ethics oversight. The company names three threats it is guarding against: compromised accounts, insiders and misuse by agents. A changelog dated September 30 clarifies that the form is an expression of interest, not an application. Organizations that handle protected health information under a business associate agreement are not covered yet.

None of this is a scientific claim. It is a policy, and its success will only be measurable later: how many organizations get in, and whether anything goes wrong. The stated goal of hundreds of organizations in the first week is a target, not a result.

Claude Science: a beta with testimonials

Claude Science is a desktop workbench for macOS and Linux, offered in beta to Pro, Max, Team and Enterprise subscribers. It bundles more than 60 skills and connectors to scientific tools, plus a reviewer agent that checks the main agent's work. Anthropic is offering up to 50 projects as much as $30,000 in credits each, and the compute provider Modal adds up to $2,000.

The evidence offered is testimonial. Manifold Bio describes its use. The Allen Institute reports about 10 reviews done with the tool. A UCSF team says a task took about a tenth of the usual time and that the results were validated independently. Those are useful signals from named institutions, but they are quotes in a launch post, without a protocol, a comparison group or a count of failures.

The enzyme: the one result you can read

The most concrete claim comes from Anthropic's own laboratory, set up in the Bay Area in spring 2026 at biosafety levels 1 and 2. The split of labor matters: Claude agents did the computational work, and humans did the bench work.

The numbers are large. About 950 agents ran for 21 hours and used 210 million tokens. They screened more than 200,000 reverse transcriptases, enzymes that copy RNA into DNA, narrowed them to 3,500 candidates and wrote 20 reports. One lead held up in the lab. Anthropic calls it ART: a reverse transcriptase from a giant phage, paired with a partner gene and an array of repeats that resembles a CRISPR array. In experiments, the array is expressed as distinct small RNAs.

Two caveats sit in the announcement itself. The function of the system is unknown, and the reverse transcriptase was already in the literature; what is new is the arrangement around it. The work is posted on bioRxiv as a preprint that has not been peer reviewed. Feng Zhang, one of the pioneers of CRISPR gene editing, is quoted calling it genuinely intriguing. That is an expert's interest, not a verdict.

Novo Nordisk: a pilot without a price

The Novo Nordisk announcement, published on GlobeNewswire on September 16, says the Danish drugmaker will test Claude Science on specific research and development workflows and will use Anthropic's models for software development. Statements from Mike Doustdar, Novo's chief executive, and Dario Amodei, Anthropic's chief executive, frame it as a strategic step. No amount, duration or financial term was disclosed. As a press release from the companies themselves, it proves the agreement exists and what it covers on paper, nothing more.

For readers who follow Anthropic's model safeguards, the biology program is a mirror image of the secrecy described in our piece on why Claude hides its reasoning: there, the company restricts what users can see; here, it relaxes what verified users can ask, in exchange for being watched after the fact. A different security question, about agents that spread on their own, is in our read of Anthropic's self replication report.

I would judge this push by its second preprint, not its first. One enzyme system of unknown function is a promising lead and a good demonstration of scale, but biology rewards the slow part: showing what a molecule does, in someone else's lab, after review. The access rules and the Novo pilot will matter only if they produce that kind of result. Until then, the honest summary is one interesting preprint surrounded by infrastructure.