Anthropic's drug development results with six partners are not one kind of evidence. Read them by type
Claude shortened a CAR T construct by about 70% in lab tests and cut a manufacturing document review by 89%. Lab results, computational output and workflow gains need different follow-up questions.
Listen to this article · 4 min
AI-generated narration of the full article.
At Enveda, Claude reviewed 1,061 pages of pharmaceutical manufacturing documentation. It caught all 11 issues the human reviewers had found and 7 they had missed, and human review time fell from more than 25 hours to under 3.
That result sits in the same Anthropic post, published October 27, as a redesigned CAR T-cell therapy component that performed better in the lab and 35,000 antibody designs generated in about a month. They are all real results reported by partners. They are not the same kind of evidence, and a life sciences company deciding where to use AI should read them separately.
Six partners, three kinds of result
| What Claude did | Reported result | Kind of evidence | |
|---|---|---|---|
| Generate Biomedicines | Proposed redesigns of a CAR T construct, with six scientists using Claude Code | About 70% shorter; roughly 10x more receptors and better tumor-cell killing in lab tests | Lab-validated, preclinical |
| BigHat Biosciences | Proposed changes to two antibody-fragment leads, connected to BigHat's lab data system | About 1,010 variants tested in 10 weeks; best binding more than 10x tighter | Lab-validated, preclinical |
| Divergence Labs | Turned research goals into a design plan and generated designs | About 35,000 antibody designs; 3 IgE-binding and 20 IgE-dislodging hits in early tests | Computational, with early testing |
| Enveda | Reviewed manufacturing documentation and drafted regulatory documents | Review time down 89%; one draft in 4 hours instead of about three months | Workflow |
| Bristol Myers Squibb | Assesses clinical trial plans for an early-stage drug | No figures; expansion planned for 2027 | Workflow, in progress |
| The Column Group | A target-assessment skill for a new company | Evaluated on 45 target-disease pairs; no outcomes reported | Computational, early |
Lab results answer a different question from workflow results
The two lab-validated results, at Generate and BigHat, are the scientifically interesting ones. Both share a structure worth copying: the model proposes, a real experiment decides, and the result goes back into the next round. At BigHat, Claude was connected to the company’s lab data system and also reanalyzed raw data, concluding that standard metrics were misleading. That is the model working inside the scientific loop, not beside it. It is also early: preclinical, in two programs, with no evidence yet about efficacy in people. Anthropic says so plainly.
The workflow results, at Enveda, are closer to what most pharmaceutical companies could use this year. Manufacturing documentation and regulatory drafting are high-volume, rule-heavy work with expensive reviewers. Here the evidence question is not “did the molecule work” but “how do we know the review was complete.” Anthropic includes the right detail: even after Claude’s review, Enveda’s standard human process found one minor error in a regulatory draft. The human step was not decoration.
What a regulated company should ask
For discovery work: what data the model was connected to, who owns the designs, and how results flow back into the next design round. The value at BigHat came from the connection to the lab system, which is an integration project, not a model choice. We saw the same pattern when Hansa partnered with Cradle on protein design.
For GxP documentation: whether AI-assisted review is part of a validated process, how its omissions are measured over time, and who signs. A reduction in review hours is only useful if the quality system can show the review remained adequate. Write the comparison method down before the pilot, as Enveda effectively did with its 11 known issues.
For access: Anthropic’s more permissive safeguards for this work are limited to vetted organizations through its Life Sciences Verification Program. Model access is part of the qualification, not a detail of the contract.
The post also mentions Anthropic’s own program: after acquiring Mithril Medicine, it says Claude helped coordinate work with 11 external labs and took a muscular dystrophy treatment idea to promising mouse results in under 18 months, with human trials possible from late 2028. It is the most ambitious claim in the post and the furthest from proof. For now, the documentation result is the one most teams can act on.