Anthropic published a progress report from its life sciences team on 27 October 2026: over recent months it has run research partnerships with six drug-developing organisations — BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group — started discovery programmes of its own, and acquired Mithril Medicine, a one-person company working on two rare forms of muscular dystrophy. The company says it will scale this work up substantially, and puts its most distant milestone at a candidate that, if results hold, could enter human clinical trials as early as late 2028. The full announcement is in Anthropic's post.
What was announced
Anthropic's argument is that drug development has no single bottleneck. The jams are spread out: picking a target, designing the molecule, optimising it so it can be manufactured reliably, writing the regulatory paperwork, planning the trials. Hence six partnerships aimed at different stages.
- Generate Biomedicines, redesigning CAR T-cell therapy. Six of the company's scientists used Claude Code to propose shorter versions of one stretch of the therapy's genetic instructions, freeing room to tackle solid tumours. One design came out around 70% shorter than the standard and, in lab tests, let immune cells build roughly ten times as many receptors, kill tumour cells more efficiently and stay active longer.
- Divergence Labs, hit generation against food allergy. The goal is an antibody with two arms: one that binds IgE and one that prises it off immune cells, rather than only neutralising what circulates in the blood. According to the company, Claude turned the objective into a design plan and generated, largely on its own, some 35,000 antibody designs in about a month. Early testing produced three candidates that bound IgE and twenty that dislodged it; three of those twenty beat the standard medication.
- BigHat Biosciences, lead optimisation. Hooked up to the company's lab data system, Claude proposed changes to two antibody fragments; the automated lab made and measured about 1,010 variants in roughly ten weeks. Midway through, the team suspected the usual heat-stability and clumping metrics were misleading, reanalysed the raw data and changed the protocol. Best results: binding more than ten times tighter, 18°C more headroom in solution for the cancer-targeting lead, and clumping over a hundred times slower for the other.
- Enveda, manufacturing and filings. Its production chemists were spending up to 75% of their time on paperwork. They connected all those documents to Claude so it could flag problems: across 1,061 pages it caught all eleven issues human reviewers had found and flagged seven more they had missed. Human review time fell by 89%, from over twenty-five hours to under three. Claude also drafted two regulatory documents; one took four hours against roughly three months by hand.
- The Column Group, company creation. The investment firm is setting up Pleiades, a company that will use Claude to run more — and riskier — programmes than a small team usually can. Its first skill, meant to judge whether a target is worth chasing, has been evaluated on 45 target-disease pairs.
- Bristol Myers Squibb, trial planning. The newest partnership, focused on assessing plans for an early-stage drug and estimating its chance of success, with expansion planned for 2027. Michael Pourdehnad, head of early clinical development in haematology, oncology and cell therapy, says:
From a single prompt, Claude produced a solid initial analysis that helped weigh different development options and would otherwise have taken days.
Separately, Anthropic bought Mithril Medicine. Its founder, Cody Tranbarger, used Claude to coordinate contracts, logistics and experiments with eleven external labs; the company says he went from idea to promising results in mice in under eighteen months.
Where it comes from
This is not the first piece of Anthropic's life sciences plan. The company had already launched Claude Science, its workbench for researchers, a credits programme for life scientists, and the Life Sciences Verification Program, which lets trusted organisations use its most capable models with less restrictive safeguards — the ones that block biological misuse and, Anthropic admits, also get in the way of legitimate research. Add to that the guest investigator call this blog covered a few weeks ago.
What it means
For anyone building software, the interesting part is not the biology but the pattern repeated across all six cases: the model is not working in an isolated chat window but plugged into the company's internal systems — the lab data manager, the manufacturing document repository — with human review that validates, corrects and sometimes rewrites the protocol.
Enveda's case travels best outside the sector: bulk document review against strict criteria, with an honest metric — how many issues it catches and how many humans missed — and review time as the savings indicator. That structure works just as well for compliance audits, tenders or contracts. BigHat's case points elsewhere: the model did not just run the plan, it questioned the metric used to judge success. And Mithril shows how far one person with delegated work gets today: eleven labs coordinated with no team behind them.
What the announcement leaves out
Almost everything published is a claim by an interested party. The figures come from Anthropic and its partners, the supporting technical reports and preprints have not been peer reviewed, and the results are mostly preclinical: no drug has been tested in people. Anthropic itself calls the work early.
Nor does it say what Mithril Medicine cost, how Pleiades is funded or how many programmes it already runs. For the 45 target-disease pairs assessed by that first skill, only the number is published, not the outcome. At Enveda, the 89% reduction was measured in a bounded test over 1,061 pages, not in sustained production. And the late-2028 date for a first human trial comes with its conditional attached: it depends on the candidates holding up.