Claude Drug Discovery Trials Cut Batch Record Review Time by 89%

What Anthropic announced

Anthropic has published its most detailed account so far of what Claude does inside working drug-development programs. The update covers research partnerships with six organizations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. Anthropic also says it has started drug-discovery work of its own.2 The company presented the results as early, and nearly all of them are preclinical. None of the candidates described has been tested in people.2325

The headline number has nothing to do with molecules. It comes from manufacturing paperwork. Enveda, which looks for drug candidates in natural sources such as plants, built a Claude-based system called the CMC Document Processor. The company tested it on 1,061 pages of batch records from four lots made for a Phase 2 manufacturing campaign. Those records were later submitted to the FDA.22 Claude caught all 11 issues that human reviewers had found and flagged seven more. Review time fell from about 25.9 hours to 2.8 hours, a cut of roughly 89%.24

Anthropic also disclosed that it had acquired Mithril Medicine. Mithril is a one-person company working on treatments for two rare forms of muscular dystrophy.28 Human trials of that program could start as early as late 2028 if the findings hold up.3

The paperwork result is the strongest result

The Enveda case is the most concrete part of the announcement because the task is clearly defined and the work is easy to check. Enveda says its manufacturing chemists had been spending up to 75% of their time preparing and reviewing documents.28 An experienced reviewer could need about ten hours to work through a single 500-page batch record from a contract manufacturer.22

The issues Claude flagged were ordinary but real. One manufacturing step ran past its permitted time limit. One test was marked as failed but had no measured value recorded.24 Enveda's quality team resolved the problems with the manufacturer before the batch was released.24 Claude also drafted regulatory documents. One first draft took about four hours, against roughly three months by hand. Human review of a second draft found one minor error.23

Outlets disagree on some details, and the differences matter. Most coverage says simply that Claude found seven issues humans missed.2729 A more detailed account says those seven were minor issues that came up in a separate, strictly timed validation run.22 Another report says Enveda's quality team judged the seven to be actionable.24 Reports also describe the human baseline differently: "more than 25 hours," "about 26 hours" or 25.9 hours.222428 These are small gaps. They still show how much the result depends on one company's dataset and one set of evaluation rules.

The system also made mistakes, and it was built with that in mind. It misread one handwritten correction but marked its reading as needing verification rather than stating it as fact.24 It runs in a view-only environment, cannot change official records, and leaves the final batch-release decision with the head of quality assurance.22

This looks like the most deployable part of the work, even if it is the least glamorous. Document review is repetitive and full of rules, and an expert can check the output quickly. Those are the conditions under which language models have been most reliable. One technical analysis notes that a production deployment would still need validated software, access controls, audit logs and accountable expert sign-off.23

Protein design: Claude runs the specialist tools

The lab results show Claude working less as a protein model and more as the operator of the field's existing computational tools. At Generate Biomedicines, six scientists used Claude Code to design part of a chimeric antigen receptor for CAR-T therapy. One design was about 70% shorter than the standard version. Cells using it produced roughly ten times as many receptors and killed tumor cells more effectively.23

Divergence Labs, which is working on food allergies, ran 22 autonomous sessions. They used about 2,750 hours of Nvidia H100 compute and produced around 35,000 antibody sequences. Structure-prediction models narrowed that pool before a few dozen candidates per arm were made and tested.22 Twenty of 91 candidate "disruptor" designs pulled already-bound IgE off its receptor. The top three outperformed omalizumab, the existing anti-IgE drug, in vitro.22 The work has been described as a preprint.

The pattern matches an earlier Anthropic study on protein binders. There, Claude designed binders against 15 targets by running publicly available design and co-folding models. Adaptyv Bio and Twist Bioscience tested the designs independently and confirmed success against 14 of the 15 targets.4 Anthropic claimed some of those binders matched or beat the best previously published designs.4

The implication for structure prediction is that the value is moving up a layer. Folding and co-folding models are increasingly treated as tools. The competitive question becomes who can chain them into a working design campaign with the least specialist labor.

BigHat paired Claude with Reccy, its automated lab platform. Its best candidates bound their targets more than ten times as tightly, one cancer candidate's thermal stability rose by 18°C, and clumping in a long-acting candidate slowed more than 100-fold.23 The figures vary by source here as well. One report gives a 6–9x affinity gain and thermal-stability gains of 8–18°C across the two antibody fragments. It also says Claude helped correct a measurement metric by reanalyzing raw data.22 Roughly 1,010 variants were made and tested in about ten weeks.

Where the coverage agrees, and the gaps

The coverage agrees on the main caveats. The results were selected and reported by the partners themselves. They use different metrics and cannot be compared on a single scale.2122 Anthropic offers no shared benchmark or independent replication.23 It also does not disclose model versions, prompts, costs, full error distributions or evaluation datasets for most of the case studies.23

The Bristol Myers Squibb work is the hardest to judge. Claude is assessing an early-stage drug's clinical-trial plans and estimating the odds that those trials succeed. The two companies plan to expand the work in 2027.2 The only published BMS quote is about speed: an analysis that would otherwise take days. No calibration data has been released, and calibration is what would matter most if the estimates were used to make decisions.23

The Column Group is setting up Pleiades, a company meant to run more drug programs, and more complex ones, than a small team usually could.2 Its first Claude tool for assessing drug targets has been evaluated on 45 target-disease pairs.3 Taken together with the Mithril acquisition, this looks like Anthropic testing the "lean biotech" idea. Mithril's founder used Claude to coordinate contracts, logistics and experiments across 11 outside labs, and reached promising mouse results in under 18 months.23

The commercial strategy

The announcement fits a product strategy built over the past year. Anthropic launched Claude for Life Sciences in October 2025 and has kept adding connectors, including ChEMBL, Open Targets, bioRxiv and Benchling.19 In June 2026 it released Claude Science, which it describes as a research workbench. CEO Dario Amodei has said he expects it to change life sciences the way Claude Code changed programming.10

The company also created a Life Sciences Verification Program. It gives vetted organizations access to versions of its safeguards that are less restrictive for legitimate biology work.8 Anthropic describes the partnerships as a feedback loop: problems found in partner labs shape model improvements, which then reach paying customers.23

Funding: Anthropic's partners in the AI protein design market

The partner list overlaps heavily with the companies drawing the most capital in AI protein design. BigHat raised a $75 million Series C in September 2026, co-led by DFJ Growth and Premji Invest.11 Generate Biomedicines raised $400 million in a March 2026 IPO.15 Elsewhere, Chai Discovery closed a $400 million Series C in July at a valuation of about $3.8 billion.1114

The totals are less broad than they look. One tracker counts $1.03 billion raised in the sector across 23 deals through late September, roughly four times the comparable 2025 figure. Without rounds above $50 million, though, capital is up only about 26% and remains below 2024 levels.13 Chai's round alone accounts for nearly 39% of the year's quantified funding.13

That concentration explains why general-purpose AI labs are moving in. Specialist protein-design startups are paying for proprietary models, wet labs and data. Anthropic is offering orchestration, operations and compliance work on top of tools many of them already use. If Enveda-style document savings can be repeated across different manufacturers and contract partners, the paperwork side of drug development may be where AI pays off first. The design gains will only prove themselves in human trials, and the earliest of those is still about two years away.3