Anthropic has spent the past several months putting Claude to work inside six organizations that make medicines, and on 27 October it published what came out of it. The partners range from antibody specialists such as BigHat Biosciences and Generate Biomedicines to Bristol Myers Squibb, Enveda, Divergence Labs, and the venture firm The Column Group. The pitch is that drug development breaks down in many places, not one, so Anthropic is testing Claude against many different jobs rather than just molecule design.

The strongest numbers come from the lab side. At Divergence Labs, which wants to treat food allergies by pulling the IgE antibody off immune cells, Claude drafted a design plan and then produced roughly 35,000 antibody candidates in about a month, largely on its own. Early tests found three designs that grip IgE and 20 that dislodge it. Of that 20, three did better than the standard medication. A combined drug has not been built yet.

Generate Biomedicines asked six of its scientists to use Claude Code on a CAR T-cell therapy, which today treats blood cancers and which Generate wants to adapt for solid tumors. One proposed design was cut by about 70% against the usual length. In lab tests, immune cells carrying it displayed around ten times the receptors, destroyed tumor cells with more efficiency, and stayed active longer.

BigHat Biosciences gives the more candid story. Over about 10 weeks its automated lab built and measured roughly 1,010 variants of two antibody fragments, with Claude proposing the changes. Midway, the team doubted the usual heat-stability and clumping measures. Claude and an Anthropic researcher reworked the analysis, BigHat refined the new protocol, and the main predictions held. The best versions bound their targets more than 10x more tightly, one stayed dissolved at a temperature 18°C higher, and the other aggregated over 100x slower.

The least glamorous result may be the one buyers notice first. Enveda says paperwork swallowed as much as three quarters of its manufacturing chemists’ hours. After connecting its batch records to Claude, the system reviewed 1,061 pages, found all 11 problems human reviewers had caught, and raised seven more they had missed. Human review time dropped 89%, with the job shrinking from more than 25 hours to under three. Claude also wrote a first draft of a regulatory document in four hours that Enveda puts at about three months by hand.

The Column Group is using Claude to set up Pleiades, a new company meant to run far more drug programs with a small team. Its first Claude skill, for judging whether a target is worth pursuing, has been checked against 45 target-disease pairs. Bristol Myers Squibb is using Claude to assess an early-stage drug’s clinical trial plan and estimate the odds it succeeds, with a wider rollout planned for 2027. Michael Pourdehnad, who heads early clinical development in hematology, oncology and cell therapy there, said that from a single prompt Claude produced an initial analysis that “would otherwise take days to complete.”

Every figure above comes from Anthropic’s own post and from technical reports its partners wrote, so none has been independently checked. Anthropic also says plainly that the work is mostly preclinical, meaning most of these candidates have not reached human testing. That is the gap that decides whether 35,000 designs or a 70% shorter construct turn into medicines, and the post gives no timeline for the partner programs reaching people.

For a lab or pharma team weighing an AI vendor, the useful signal sits in the Enveda and BMS examples: document review and trial planning can be tested this quarter, while a design win only counts once it survives a human trial.

Reported by Anthropic’s research blog on 27 October 2026.