What the reports behind Anthropic's drug-development claims actually show

On October 27 Anthropic published a progress report on its work with six drug developers, each case summarized in a few sentences and linked to a partner's technical report. Two of those reports are long enough to check the summaries against: Divergence Labs' antibody preprint and Enveda's paper on batch-record review. The numbers in Anthropic's post match them. What the post leaves out is mostly the fine print, and in both cases that fine print changes how strong the result is.

Divergence Labs: better than omalizumab, not better than the starting molecule

Anthropic's version: Claude helped generate roughly 35,000 antibody designs in about a month; Divergence found three that bind IgE and 20 that dislodge it from its receptor, and "three of those 20 appeared to dislodge IgE more than the standard medication."

The preprint gives the funnel. The disruptor arm started from 11,273 designs, 91 were tested, and 20 showed activity in a competitive bio-layer interferometry (BLI) screen. For the anchor arm, 23,600 designs produced 57 that were tested, and 3 bound IgE, with apparent affinities of 3.4–235 nM. In a dose-titration assay three purified variants, V1, V4 and V8, "removed more IgE than omalizumab at every concentration tested" (25–400 nM).

The comparison that matters more is the one against sdAb026, the published single-domain antibody that every disruptor design was a redesign of. Six variants beat it in the initial screen. Once eight active designs were purified and titrated, the authors write, "the variants did not appear to improve upon the parent sdAb026." In the discussion they repeat it: "we did not measure any variants with improved functional activity over sdAb026 at controlled doses." What they claim instead is diversification: 20 working variants up to eight substitutions away from the parent, which give several starting points for tuning developability and other properties.

The authors call the dependence on an existing sequence, and on its docked structure, "a key limitation" of the method. Everything so far is purified protein against an immobilized receptor fragment, and the designs are monovalent. Basophil and mast-cell assays, the biparatopic molecule itself and animal studies are listed as next steps. The abstract also says the campaign was "guided by our research team": Claude ran the literature review, model selection and design phases, with humans steering.

Enveda: "all 11" across runs, not in a single pass

Anthropic's version: on 1,061 pages of documentation, "Claude caught all 11 issues that human reviewers had spotted—and flagged seven they'd missed," cutting review time by 89%, from over 25 hours to under three.

The Enveda paper describes CDP, a system that runs Claude over scanned, mostly handwritten executed batch records, here four from one Phase 2 GMP campaign. The time figures hold: 166 minutes for CDP processing plus the reviewers' check of its comments, against 1,553 minutes (25.9 hours) of unaided manual review, a 9.4-fold difference. Model calls cost US$36.97 at list prices, about 3.5 cents a page.

The detection claim is narrower than the summary makes it sound:

• CDP raised each of the 11 human-found issues "in at least one of its runs," so the paper does not say a single run caught all 11. For two of them the adjudicator had to interpret: CDP noted non-reconciling charge tables without calling them discrepancies, and it flagged a temperature spike without calling it an excursion. Both were counted as matches.
• The seven extra findings came from one run on Claude Opus 4.7. All seven were verified as accurate, and CDP itself graded two as Minor and five as Informational: an 8-minute overrun on a rinse soak, a yield one point above range, a set point rewritten without confirmation. A rerun on Opus 5.5 found six additions, four of them new, which shows how much the output varies from run to run.
• Each run misread one handwritten entry and posted it at Major.
• About four in five comments needed no action, for humans (79 of 96) and for CDP (33 of 42) alike.
• Human reviewers could consult earlier-step records that CDP was not given.

One adjudicator scored every comment and a second person confirmed the labels. The authors are Enveda employees, call the analysis exploratory, and say CDP "is not validated as a GMP computerized system" and is used only as an advisory aid. They also note the development runs kept a fallback model switched on, which future validation will turn off. Their own conclusion is more modest than Anthropic's: "In this study CDP added to a human review and did not replace one." Enveda released a generic companion implementation of the deterministic parts, open-doc-review, under the MIT license. The production prompts and rules stay proprietary.

The rest

Generate Biomedicines' work is now a bioRxiv preprint. It describes a 45-amino-acid CAR signaling domain, less than a third the length of the clinical 4-1BB/CD3ζ domain, with higher surface expression and better B-cell depletion in its tests. We did not check the full text against Anthropic's "roughly 10 times as many receptors" figure. BigHat's report page loads its content only through scripts, and we could not read it. The BMS, Column Group and Mithril programs are described only in Anthropic's own words, plus a technical report on the in-house LAMA2 muscular dystrophy program.

None of the results has been independently replicated, and every report comes from a company with a stake in it. Within those limits, the two reports we could check back up the headline numbers. They do not support reading them as "Claude beats the existing drug" or "Claude replaces the reviewer." Divergence's designs match the antibody they were derived from, and they beat omalizumab only in an in vitro assay. Enveda's system is a fast first pass that a human still has to check.