Anthropic says Claude is now doing real work inside six drug-development programs, and that it has started its own. In an October 27, 2026 research post, the company reports partner results with BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group, and discloses that it acquired Mithril Medicine, a one-person company working on rare childhood muscular dystrophies.
The honest framing comes from Anthropic itself: "all of this research is still at an early stage," and "so far it is mostly preclinical." Nothing here is a drug approved or even dosed in a person. What is new is the breadth, spanning molecule design, lab optimization, manufacturing paperwork, company creation and trial planning, plus numbers a reader can inspect.
TL;DR: the six partners and one acquisition
| Partner | Stage | What Claude did | Reported result |
|---|---|---|---|
| Generate Biomedicines | Modality engineering | Six scientists used Claude Code to propose shorter CAR T genetic designs | One design about 70% shorter; roughly 10x the receptors, better tumor killing, longer activity in lab tests |
| Divergence Labs | Hit generation | Turned goals into a design plan, then generated about 35,000 antibody designs in a month | 3 designs bound IgE; 20 dislodged it; 3 of those beat the standard medication in early tests |
| BigHat Biosciences | Lead optimization | Proposed edits to two leads, connected to the lab data system | About 1,010 variants in about 10 weeks; best binders over 10x tighter |
| Enveda | Manufacturing and filings | Reviewed batch records, drafted regulatory documents | Caught all 11 known issues plus 7 missed; review time down 89% |
| The Column Group | Company creation | First skill for judging drug targets, for new company Pleiades | Skill evaluated on 45 target-disease pairs |
| Bristol Myers Squibb | Clinical trial planning | Assessed an early-stage trial plan and its odds | Early work; expansion planned for 2027 |
| Mithril Medicine (acquired) | Anthropic's own program | Ran operations and experiments across 11 external labs | Idea to promising mouse results in under 18 months |
What did Anthropic actually announce?
Three things at once. First, a progress report on six partnerships that began "over the past several months." Second, technical reports and preprints from the partners describing methods and early results. Third, a statement of intent: Anthropic says it plans "to scale this work up substantially from here."
The strategic argument is that drug development has no single bottleneck. Anthropic lists the chain: target identification, molecule design, manufacturing optimization, regulatory filings, trial design and administration. Its bet is that AI involved across all of them beats AI that only proposes new molecules. It also says the programs let it prioritize specific model improvements based on where Claude struggled.
This continues a run of Anthropic life-science moves we have tracked: Claude Science, the AI workbench for scientists, the Life Sciences Verification Program, the Novo Nordisk collaboration, rare disease research grants and the guest investigator program. The October 27 post is the first to show outcomes across many stages together.
Generate Biomedicines: a CAR T design 70% shorter
CAR T-cell therapy has helped with blood cancers but struggles with solid tumors. Generate wants to shorten one part of the therapy's genetic instructions to free room for extra instructions that suit other cancers.
Six Generate scientists used Claude Code to propose new designs for that part. Anthropic reports one was about 70% shorter than the standard version. In lab tests it let immune cells make roughly 10 times as many receptors, killed tumor cells more efficiently and stayed active longer.
The caveat is the usual one: lab tests in cells are not patients. A shorter construct that performs well in a dish still has to survive animal studies and trials. The notable part is the working pattern, scientists driving a coding agent rather than a specialized biology model.
Divergence Labs: 35,000 antibody designs in a month
Divergence is working on a food allergy approach. Instead of only neutralizing free IgE in blood, as standard medications do, it wants a drug that detaches IgE from immune cells. That needs a two-armed molecule: one arm that binds IgE tightly and one that helps pry it off.
Per Anthropic, Claude turned the goals into a design plan and then, "largely autonomously," generated roughly 35,000 antibody designs in about a month. Early testing found three designs that bound IgE and 20 that dislodged it, and three of those 20 appeared to dislodge IgE more than the standard medication. The next step is building a single drug combining both arms, which has not been shown yet.
Read the funnel honestly: 35,000 designs produced 23 hits across two separate functions. That hit rate is what a design-and-screen pipeline looks like, and the value is the speed of getting to a testable list.
Protein structure ribbon illustrating antibody design for Anthropic drug development
BigHat: Claude questioned the lab's own metrics
BigHat pairs AI with an automated lab. Scientists and an Anthropic researcher worked with Claude, connected to BigHat's lab data system, to improve binding, stability and clumping in two antibody-fragment leads. The lab made and tested about 1,010 variants in about 10 weeks.
The most interesting detail is not the improvement but the correction. Partway through, the team suspected the standard heat-stability and clumping metrics were misleading. Claude and the researcher reanalyzed the raw data, proposed different metrics and a protocol change, BigHat's scientists refined and ran it, and the main predictions held.
Final numbers versus the originals: the best versions bound more than 10x more tightly, the cancer-targeting lead stayed in solution 18°C higher, and the longer-lasting-drug lead clumped more than 100x more slowly. That is evidence of an agent doing analysis, not just generation, though it was a single researcher-supervised collaboration.
Enveda: where the biggest time savings were
At Enveda, manufacturing chemists spent up to 75% of their time on paperwork. Enveda connected its manufacturing documents to Claude and had it flag possible issues.
Clinical documentation review illustrating Claude cutting Anthropic drug development paperwork
Across 1,061 pages, Claude caught all 11 issues human reviewers had found and flagged seven they had missed. Review time fell 89%, from more than 25 hours to under three. Separately, Claude drafted two regulatory documents. One took four hours versus about three months by hand; for the other, Enveda's standard human review found one minor error.
This is arguably the most transferable result. It needs no new science, only a model that reads long documents accurately, and the bar is checkable: known issues found, new ones flagged. Note that "seven they'd missed" is Enveda's assessment, and a missed issue that matters is the dangerous failure, so humans stay in the loop.
The Column Group, Pleiades and Bristol Myers Squibb
TCG is establishing Pleiades, a company designed to use Claude to run many more, and more complex, drug programs than a small team normally could. Its first Claude skill, which helps judge whether targets are worth pursuing, has been evaluated on 45 target-disease pairs. Anthropic frames this as a possible template for cutting both scientific and administrative overhead so investors back harder ideas.
Bristol Myers Squibb is the newest partner. It uses Claude to assess an early-stage drug's trial plan and estimate the likelihood of success. BMS's Michael Pourdehnad said Claude generated "a robust initial analysis" from a single prompt that "would otherwise take days." The two plan to expand in 2027. Success probabilities are a forecast, so the real test comes when trials read out.
Mithril Medicine: Anthropic's own drug program
The acquisition is the part that changes Anthropic's posture from vendor to operator. Mithril was a one-person company developing treatments for two rare forms of muscular dystrophy. Founder Cody Tranbarger used Claude to organize contracts, logistics and experiments with 11 external labs and delegated much of the company's operations to it.
Anthropic says that took a possible treatment for one form from idea to promising mouse results in under 18 months. Tranbarger is now at Anthropic and plans several new programs. If candidates keep holding up, Anthropic expects one could start human trials as soon as late 2028. That date is a company expectation, not a commitment.
What is verified, what is not
| Claim | Status |
|---|---|
| Six named partnerships exist | Named by Anthropic; partners published technical reports |
| Lab and document numbers | Reported by Anthropic and partners; we have not independently reproduced them |
| Any drug works in humans | Not shown; work is mostly preclinical |
| Mithril human trials by late 2028 | Anthropic's expectation, conditional on results holding |
| Claude proposes molecules better than specialists | Not claimed; humans test, select and refine |
Compare this with OpenAI's acquisition of Orchestra Bio, which also puts agents inside drug development, and with Anthropic's Claude scorpion toxin binder work. The pattern across labs is the same: general-purpose agents plus real wet-lab partners, with the lab doing the validating.
Why it matters, and the risks
For builders, the signal is that agentic coding tools are being pointed at regulated, high-stakes workflows with measurable checks. The Enveda and BigHat results are the replicable lessons: connect the agent to the real data system, define what a correct answer is, and let it challenge assumptions.
For everyone else, the caution is selection. Case studies in a company's own post are the successes. Anthropic does not publish how many partner experiments failed, and drug development has a long record of promising preclinical results failing in people. Anthropic also notes that safeguards against misuse can impede legitimate biology work, which is why it created the verification program.
If you work on a biotech or pharma team, the practical next steps are Claude Science, Claude credits for life scientists and the open call for guest investigators. If you are evaluating agents inside regulated workflows, our agent-safety coverage and AgentBeam, the agent security platform from the explainx.ai team, are relevant to controlling what such agents can touch.
Related reading
- Claude Science: the AI workbench for scientists
- Anthropic Life Sciences Verification Program
- Anthropic Guest Investigator Program
- Novo Nordisk and Claude Science
- OpenAI acquires Orchestra Bio
- Claude scorpion toxin protein binders
- Anthropic rare disease research grants
Details are accurate as of October 27, 2026, based on Anthropic's published post; partner technical reports are linked from it.