Anthropic published an update on October 27, 2026 on how Claude was used by six organizations developing new medicines, from antibody design to the paperwork that keeps a trial from starting. It is a case series, not a product launch: each partner contributed a technical report on what it asked and what came back. The sharpest number in it is not about molecules. At Enveda, a regulatory document that takes about three months to write by hand was drafted in four hours, and human review of manufacturing records went from more than 25 hours to less than three.
What was actually released
The post is published on Anthropic’s research site, dated October 27, 2026, and names six partners: Generate Biomedicines, Divergence Labs, BigHat Biosciences, Enveda, The Column Group and Bristol Myers Squibb. Anthropic also disclosed the acquisition of Mithril Medicine, a one-person company working on two rare forms of muscular dystrophy, and says it now runs drug programs of its own.
| Partner | What Claude was asked to do | Number published |
|---|---|---|
| Generate Biomedicines | Propose designs for a shorter CAR T receptor sequence | One design about 70 percent shorter, roughly 10 times more receptors in lab tests |
| Divergence Labs | Generate antibody designs for a food allergy drug | About 35,000 designs in a month, 23 early hits |
| BigHat Biosciences | Improve binding, stability and clumping in two antibody fragments | About 1,010 variants in 10 weeks; best bound more than 10 times tighter |
| Enveda | Check manufacturing records, draft regulatory documents | 1,061 pages checked, all 11 known issues found, review time down 89 percent |
| The Column Group | Build a drug company run on Claude | First Claude skill evaluated on 45 target-disease pairs |
| Bristol Myers Squibb | Assess trial plans and estimate probability of success | No figure published |
The Enveda case is the most legible because it has a control. Chemists there spent up to 75 percent of their time on documentation, so the company connected its manufacturing records to Claude and ran it against reviewers who had already found 11 issues. Claude found all 11 and flagged seven more that the humans had missed, and review time fell from over 25 hours to under three.
What the numbers actually cover
- Six partners, each with its own technical report, plus Anthropic’s own programs
- One measured result against a human baseline: Enveda’s review of 1,061 pages
- One regulatory draft produced in four hours against about three months by hand
- Every case is preclinical, and no program has reached human trials
Divergence Labs is the case that reads most like a discovery story. Claude turned the company’s design goals into a plan and then produced about 35,000 antibody designs in roughly a month, largely without supervision. Tests found three designs that bind IgE and 20 that dislodge it from immune cells; three of the 20 appeared to beat the standard medication in early tests.
What the claim rests on
Every number here is reported by the partner that ran the work, and Anthropic chose which partners appear. The post is closer to a portfolio of testimonials with appendices than to a study, and the parties that benefit wrote the appendices.
The strongest claim is the Enveda one, precisely because a human reviewer had already produced the ground truth. The weak point is the seven extra flags: the post does not say whether those seven were confirmed as real issues, so a reader cannot tell them apart from well-formed false positives. A review that misses nothing and adds noise is a different product, and 1,061 pages is a single test.
The BigHat numbers have the same shape: about 1,010 variants made and tested over 10 weeks, with the best binding more than 10 times more tightly, staying in solution 18 degrees Celsius higher and clumping more than 100 times more slowly. Those are the best variants, not the average.
The Divergence count invites the wrong comparison. Thirty-five thousand designs is a throughput figure, useful only next to a conventional baseline that is not published. Twenty-three early hits out of that many designs is a hit rate near 0.07 percent, and nothing here says what a well-run antibody campaign normally produces.
What Anthropic does not publish
Which model each partner used beyond “Claude Code” and “our most powerful models”, what the work cost in tokens or dollars, how many of the 35,000 Divergence designs were actually synthesized, whether the seven extra Enveda flags were confirmed, and any independent replication.
All of this research is still at an early stage. In particular, so far it is mostly preclinical.
That sentence is Anthropic’s own, and it is the fairest summary of the release.
On this page
What it changes for teams buying AI for regulated work
Nobody here shows a model inventing a better molecule than a chemist would. The gains land in document review, filing drafts, variant triage and trial planning, the parts of drug development that consume calendar time without producing a patent.
That changes which buyer should care: a biotech evaluating Claude for research finds anecdotes and no benchmarks. An operations or regulatory team has one number with a control, an 89 percent cut in review time, plus a drafting result that compresses a three-month task into four hours. Those are workflows you can test this month with documents you already own, and the failure mode is visible, because you know which issues are in the pile.
Access is administrative. Anthropic offers Claude credits for life scientists, opened Claude Science as a research workbench, and runs a Life Sciences Verification Program that gives trusted organizations more permissive safeguards. Those are the safeguards a general account keeps, so the useful configuration sits behind an application.
The Neural Minute take
Our own record on Claude in science is one well-documented case: in Anthropic’s October 8 report on Claude Science, an astrophysicist used the workbench to fill the unobserved third of the ultraviolet sky, and the value came from doing a repetitive, checkable task faster. This post says the same thing at a larger scale.
For anyone paying for AI tools, the read is that the industry’s drug story has moved from discovery to paperwork, and paperwork is where the headcount is. Regulated documentation is a better first target than molecule design: the inputs are yours, the ground truth exists, and the job is to be complete rather than creative.
Two caveats keep this from being a buying signal. The results come from partners Anthropic chose, at a moment when the company is selling an IPO story and needs life sciences to look real. And the number that would settle the argument, the cost of these workflows at a normal company’s volume, is the one missing from every case. A team that wants to know whether Claude pays for itself should reproduce the Enveda test on its own documents, count confirmed issues rather than flags, and price the hours it saves.
Sources
- Early progress on accelerating drug development with Claude, Anthropic, published October 27, 2026, read on October 28, 2026: https://www.anthropic.com/research/early-progress-drug-development
- The date of the fact is that page’s publication date, not the age of the feed item that surfaced it
- The review figures, the four-hour draft and the BigHat numbers are as written on that page, which links each partner’s report
- The preclinical caveat is a direct quote from Anthropic’s own section on extending the work