Anthropic published an update on October 27 describing months of drug-development work with six organizations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group. The company also disclosed that it has started running its own discovery programs and acquired Mithril Medicine, a one-person company founded by Cody Tranbarger that used Claude to coordinate 11 external labs and took a muscular dystrophy treatment from idea to mouse results in under 18 months.
The framing in the post is the interesting part. Anthropic’s science team argues there is “no single problem to solve” and that “the bottlenecks are distributed all throughout the drug development process.” That is a different bet from the one most AI-for-biology companies have made. It says the value is not in a better molecule generator but in an agent that can sit inside the tedious middle of the pipeline: the optimization loops, the batch records, the trial plans, the regulatory drafts.
The numbers are the argument
The case studies carry unusually specific results, and the specificity is what makes them worth reading.
At Generate Biomedicines, six scientists used Claude Code to redesign part of a CAR T-cell therapy’s genetic instructions. One design came out about 70% shorter than the standard version, and in lab tests it let immune cells produce roughly 10 times as many receptors, killed tumor cells more efficiently, and stayed active longer. At Divergence Labs, Claude converted goals for a food-allergy drug into a design plan and then generated roughly 35,000 antibody designs in about a month; early tests found three binders and 20 designs that could dislodge IgE from immune cells, three of which beat the standard medication at that job.
BigHat Biosciences is the most instructive case because it shows the model doing science rather than filling in a template. Claude, connected to BigHat’s lab data system Reccy, proposed changes to two antibody fragments, and the lab made and tested about 1,010 variants in 10 weeks. Partway through, the team suspected the standard heat-stability and clumping metrics were misleading. Claude and an Anthropic researcher reanalyzed the raw data, proposed different metrics and a protocol change, and the main predictions held. The best variants bound more than 10x more tightly, one stayed in solution 18°C higher, and another clumped more than 100x more slowly.
At Enveda, the result is less glamorous and possibly more consequential. Manufacturing chemists spent up to 75% of their time on paperwork. A system connecting manufacturing documents to Claude flagged every one of the 11 issues human reviewers caught across 1,061 pages, plus seven they missed, and cut human review time by 89%, from over 25 hours to under three. Claude also drafted two regulatory documents; one took four hours against roughly three months by hand.
What this actually demonstrates
Read the whole post and a pattern shows up. The wins cluster in work that is text-heavy, iterative, and checkable. Antibody design proposals get tested in a wet lab. Manufacturing records get compared against a standard. A regulatory draft gets reviewed by a human who found one minor error. In each case there is a cheap verification step, and Claude is doing the generation while something else does the judging.
That is a real capability and also a bounded one. The post says so itself: “All of this research is still at an early stage. In particular, so far it is mostly preclinical.” No drug described here is in humans. The Bristol Myers Squibb work is trial planning and success-probability estimation, with expansion planned for 2027. Michael Pourdehnad, BMS senior vice president and head of early clinical development for hematology, oncology and cell therapy, is quoted saying a single prompt produced an initial analysis that “would otherwise take days to complete.” Days, not months. The gap between the headline numbers and the clinical reality is where the skepticism belongs.
The wins cluster in work that is text-heavy, iterative, and checkable. Claude generates; something else judges.
The Mithril acquisition deserves its own scrutiny. Anthropic bought a company with one employee, and the pitch is that Claude let that one person run experiments across 11 external labs. If true, it is a striking data point about operational leverage. It is also a sample size of one, run by a founder who now works at Anthropic, with a drug candidate that Anthropic itself says might reach human trials “as soon as late 2028.” That is a long way off and a conditional.
The part that matters for AI builders
The strategic move here is the vertical integration. Anthropic is not just selling tokens to biotech. It is running its own programs, acquiring assets, building a research workbench called Claude Science, offering credits to life scientists, and running a Life Sciences Verification Program that gives trusted organizations “more permissive versions of our safeguards.” That last item is the one to watch. Anthropic is arguing that its biosecurity guardrails impede legitimate research and that the fix is a vetted-access tier. Reasonable, and also a governance question the whole industry will have to answer in public.
There is a model-improvement loop buried in the post too. Anthropic says the partnerships let it understand “the strengths and limitations of our models” and that it has “already prioritized specific improvements to our models based on what we’ve learned so far.” Deep partnerships as a data flywheel for the next model. The drug programs are the product and the training signal at once.
What to watch is not the next flashy design number. It is whether any of these programs clears an IND and enters Phase 1, and whether the Enveda-style paperwork gains hold up under an actual regulator’s review rather than an internal one. Anthropic says it plans to scale this work “substantially.” The next update should be judged on how many of these candidates are still alive.