Anthropic published a post on 27 October 2026 describing early results from partnerships with six drug developers. Two of those partners released their own write-ups, which lets us do something press coverage rarely does: compare the summary with the underlying documents. This article reads AI drug discovery claims at three levels: what each partner measured, what the summary implies, and how far any of it sits from a medicine a patient could take.[1]
What Anthropic reported
The post covers six partnerships, each at a different stage of the pipeline: Generate Biomedicines (engineering a CAR T-cell component), Divergence Labs (antibody design for food allergy), BigHat Biosciences (optimising antibody fragments), Enveda (manufacturing documents and regulatory filings), The Column Group (founding a company, Pleiades, around Claude-assisted target assessment) and Bristol Myers Squibb (clinical-trial planning). It also describes Anthropic's acquisition of Mithril Medicine, a one-person company working on two rare forms of muscular dystrophy.[1]
The headline figures in the post are specific: about 35,000 antibody designs generated in roughly a month for Divergence; a 1,061-page document review at Enveda; about 1,010 variants made and tested over roughly 10 weeks at BigHat. We could check two of these against the partners' own documents. We could not check the rest, and we say so below.
Case 1: Divergence Labs, where the numbers hold but the baseline matters
The Divergence preprint is a long technical document with methods, assay conditions and a limitations discussion. It confirms the post's headline numbers. Claude-orchestrated campaigns produced 11,273 redesigns of a known disruptor antibody and 23,600 de novo designs for a second "anchor" arm; 91 disruptor designs and 57 anchor designs were nominated and tested in the lab.[2]
Three details change how the result reads:
- The starting point was an existing antibody. The disruptor designs start from sdAb026, a single-domain antibody with previously reported disruptive activity. The authors write that they "did not measure any variants with improved functional activity over sdAb026 at controlled doses". What the campaign produced was 20 active variants of a known molecule, a faster route to diversity, not a better molecule.[2]
- The comparison with the approved drug is narrow. Three variants removed more IgE than omalizumab at the concentrations tested, in a bio-layer interferometry assay run in duplicate with purified proteins. That is a binding experiment, not cells, not animals, not people. The authors list basophil and mast-cell assays, allergen challenge and animal studies as the work still to come, and name anaphylaxis risk from receptor cross-linking as a key risk.[2]
- The de novo half was harder. Only 3 of 57 anchor designs bound IgE, a hit rate the authors put at about 5%. No design aimed at the consensus epitope bound, and the authors note that the critic scores did not reliably predict which designs would bind.[2]
Our interpretation: this is a credible demonstration that an agentic design loop can generate many candidates and that a minority survive contact with a binding assay. It is not evidence that a better allergy drug exists. The preprint itself says the pieces are "the starting point for further characterization".
Case 2: Enveda, a documentation task with honest caveats
Enveda's study is the most checkable item in the post because it is a document-review task with a human baseline. The authors describe a system that sends batch records to Claude in windows of pages, with a reconciliation pass, and posts advisory comments for a named person to accept or reject. It was tested on four executed batch records (1,061 pages) from one Phase 2 campaign.[3]
The reported results match the post: human reviewers raised 11 actionable issues and the system raised all 11 in at least one run, plus 7 further actionable issues that no human raised and that an adjudicator verified. Review took 166 minutes with the tool and a quality check versus 1,553 minutes unaided, a 9.4-fold reduction, which is consistent with the post's "89%" figure and its "over 25 hours to under 3".[3]
The authors then list the caveats themselves:
- The scores are "descriptive" and the analysis "exploratory".
- Most comments on both sides needed no action (82% of human comments, 79% of the system's).
- Each run contained a misread, including a struck-through handwritten correction.
- The tool "is not validated as a GMP computerized system and is used as an advisory aid".
- One consultancy's time had no log, and the AI arm worked from pre-flagged errors, so the time saving reflects that presentation.
- All authors are Enveda employees who built the tool for their own use.[3]
None of this undermines the result. It defines it: a promising small study of a back-office task, where the cost is dollars per document and a human still decides.
What we could not check
The BigHat report page linked from the post did not render readable text when we fetched it, so we have not verified the BigHat figures (more than 10x tighter binding, 18°C higher stability, more than 100x slower clumping) and report them only as Anthropic's summary. We did not obtain readable text from Generate Biomedicines' publication either, so the 70% shorter design and tenfold receptor figures are likewise unverified here. The Bristol Myers Squibb item is a quoted statement of an analysis, not a measured result, and the Pleiades item reports an evaluation on 45 target-disease pairs without results.[1]
How far is this from a medicine?
The FDA's clinical-research page gives approximate rates of progression: about 70% of drugs move on from Phase 1, about 33% from Phase 2 and about 25-30% from Phase 3.[4] Everything above happens before Phase 1. Our reading is that a lab win at the design stage tells you a method can produce candidates, not that any candidate will clear those later gates. Anthropic's post reflects this: for Mithril, it says a drug could be ready for human trials "as soon as late 2028" if candidates hold up.[1]
A checklist for any AI drug-discovery claim
- Name the stage. Design, binding assay, cells, animals, Phase 1, 2 or 3. Each is a different claim.
- Name the baseline. Compared with a known parent molecule, an approved drug, a human reviewer or nothing?
- Check the denominator. 3 of 57 and 20 of 91 are different stories from "35,000 designs".
- Find who wrote it. Company-authored write-ups are useful primary sources and also interested parties.
- Look for the authors' own limitations. In both documents we read, they were stated plainly. If a summary drops them, go to the source.
- Separate speed from quality. Faster iteration, fewer person-hours and a better molecule are three different outcomes.
For how this fits the broader habit of reading vendor claims, see How to Read an AI Benchmark Claim Without Being Misled and the research topic hub.
