Anthropic's drug discovery reports: what Claude did, and how each lab recorded it

On 27 October 2026, Anthropic described its drug development work with Claude: research partnerships with six organizations (BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group) and a drug discovery program of its own, Mithril Medicine, which it acquired in July. The post reports a much shorter signalling domain for CAR T-cell therapy, antibodies that pull IgE off immune cells, faster quality review of manufacturing records and a muscular dystrophy antibody that worked in mice.

Anthropic is careful about the stage: "All of this research is still at an early stage," and "so far it is mostly preclinical". None of the drugs has been tested in people.

The post links technical reports for five of the seven programs. We read them with a different question from the headline numbers: what exactly did Claude do, what did people decide, and how does each report record the difference?

Each report answers it in its own way. One lists which model proposed which design. One draws which phases Claude ran nearly on its own. One logs every model call in a regulated system. One counts a week of emails and warns that its own account of the AI's work may be biased.

Below: what each program reported, what Claude did in each, how each report records it, what the human checks caught and missed, the four formats side by side, and what a shared record could add. Every source was checked on 27 October 2026. These are company reports, not peer-reviewed papers, and we do not assess the science in them.

What did Anthropic announce?#

Seven programs at different stages of drug development. In the words of Anthropic's post:

• Generate Biomedicines: CAR T-cell therapy. "Six Generate scientists used Claude Code to propose new designs" for the signalling part of a chimeric antigen receptor (CAR), the receptor that lets engineered immune cells find cancer. One design was about 70% shorter than the standard version and, in lab tests, let cells make roughly 10 times as many receptors.
• Divergence Labs: food allergy. "Claude converted Divergence Labs' goals into a detailed design plan, and then, largely autonomously, generated roughly 35,000 antibody designs in about a month to be narrowed down and tested." Testing found three designs that bind IgE and 20 that dislodge it.
• BigHat Biosciences: improving drug candidates. Claude, connected to BigHat's lab data system, proposed changes to two antibody fragments, and BigHat's lab made and tested about 1,010 variants in about 10 weeks. The best versions bound more than 10 times more tightly than the originals.
• Enveda: manufacturing records. On 1,061 pages of batch records, "Claude caught all 11 issues that human reviewers had spotted—and flagged seven they'd missed," and human review time fell by 89%.
• The Column Group: company creation. TCG is founding Pleiades, a company that will use Claude to run more drug programs than a small team usually can. Its first Claude skill, for judging whether a drug target is worth pursuing, "has been evaluated on 45 target-disease pairs."
• Bristol Myers Squibb: clinical trial planning. BMS uses Claude "to assess plans of an early-stage drug in clinical trials, and to estimate the likelihood of those trials succeeding." The two companies plan to expand the work in 2027.
• Mithril Medicine: Anthropic's own program. Its founder, Cody Tranbarger, "used Claude to organize contracts, logistics, and experiments with 11 external labs." The program went from an idea to promising results in mice in less than 18 months. If the candidates hold up, Anthropic expects a drug to be ready for human trials "as soon as late 2028."

The Column Group and BMS sections link no technical report. The other five do, and the rest of this post is about them.

What did Claude do in each program?#

The reports describe the division of work in more detail than the announcement. In each report's own terms:

In all five, people set the goals, and the lab work was done by the companies' own labs or the contractors they hired. The interesting differences are in how each report writes down the rest.

How does each report record Claude's part?#

Generate: which model proposed which design

Generate's report, "Refactored intracellular domains enhance CAR expression and function", is the only one that traces the source of a design to a named model. In round 1 the team "used Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.2, Kimi K2.5 and MiniMax M2.5", models from four companies. The methods name the origin of an early design that later work built on: "brs–ζ1–40 was proposed by Claude Opus 4.6 in round 1." Rounds 2 and 3 used Claude Code, running Opus 4.7 and Opus 4.8.

The designs are filed by idea, "a design question with a written rationale". The methods say: "Each design was assigned to one idea when it was proposed, and the agents gave each idea a short code name." Recording the idea at the moment of proposal, rather than afterwards, is what lets the report test which ideas paid off. It goes on to evaluate the agents' design hypotheses against the screens.

The supplement shows the limits of keeping such a record by hand. "We recovered 17 distinct prompts from 34 archived round 1 agent runs," it says, and the table that follows lists "selected model assignments". Some prompts went to one model and some to as many as four. That is honest bookkeeping, and it also shows that the prompt record had to be recovered from archives.

Generate also built its own electronic lab notebook, with Claude Code as "the primary interface". After each experiment, "agents use plate maps to link measurements to designed variants and experimental conditions for analysis and plotting." The report has no separate AI-use statement: the agents' role is described in the methods. Its lead design, krgr–ζ1–40, is 45 amino acids long, "less than a third the length" of the 4-1BB–CD3ζ domain used in clinical CAR T-cell therapy.

Divergence: phases Claude ran, and phases with people

Divergence's preprint, "Towards a best-in-class IgE disruptor for food allergy with Claude-enabled antibody design", names the model as Claude Mythos 5.1 and says the campaign "was largely driven by interaction with" it. First, "in conjunction with our research team, Claude orchestrated high-level design planning and decomposed the stated objective into distinct work phases." Then it "proceeded with near-autonomous execution of these work phases, supported by our research team throughout."

Its Figure 3 is a map of who did what, drawn as a diagram: "blue boxes are near-autonomous Claude phase execution, green frames Claude-driven objective decomposition with human interaction." In total, "the campaign consisted of a total of 22 distinct phases of autonomous execution, of which 15 were primarily devoted toward computational design."

That is a contribution record at the level of work phases. The abstract gives the human role in one phrase, "Guided by our research team", and the acknowledgements thank people only. There is no separate AI-use statement. The 35,000 designs are Anthropic's figure. The preprint reports the lab side: of 91 disruptor designs characterized experimentally, 20 were active, and three removed more IgE than omalizumab, the approved antibody, at 25–400 nM.

Enveda: every model call in an append-only log

Enveda's report, "CDP: a language-model system for quality review of executed pharmaceutical batch records", is the most formal about records, because its setting demands it. Drug manufacturing falls under Good Manufacturing Practice (GMP), and the document system CDP reads is regulated under 21 CFR Part 11, the FDA's rule for electronic records.

CDP's controls, in the report's words:

• "Every AI model call made in review, extraction and data capture is logged with the AI model, the instructions, the response and the token use, in an append-only log stored in the ECM."
• "There is no write path from the AI model to the system of record."
• "A named person, working under their own account, decides on every comment and takes every consequential action, such as accepting a record, sending it back to the manufacturer or releasing a batch."

Together, the report says, these "allow any review to be reconstructed afterwards: what CDP was given, which version of its instructions it used, what it returned and what a person decided."

It is just as direct about what CDP is not. "CDP is not validated as a GMP computerized system and is used as an advisory aid." Enveda has no zero-data-retention agreement with the model provider. The report notes that draft EU GMP Annex 22 "would exclude large language models from critical GMP applications" and designs CDP as a non-critical, advisory tool. Models are named per run: Claude Opus 4.7, and a second run on Claude Opus 5.5, which raised six actionable comments no human reviewer had raised, four of them not found in the earlier runs.

Anthropic's post adds that Claude wrote first drafts of two regulatory documents for Enveda, one in four hours against about three months by hand. The CDP report does not cover those drafts.

Mithril: a week counted, and a statement that names its own bias

Mithril's report, "Preclinical Characterization of a Bispecific Bridging Antibody for the Treatment of LAMA2-Related Congenital Muscular Dystrophy", is dated 26 October and marked "Not peer reviewed". Besides the antibody results, it has a full section on how the team works with Claude, built around one week: "In one week, one of us, working with Claude, turned a 341-page data package on 113 constructs into one plan across five CROs", the contract research organizations that ran the experiments.

The setup: "Each task has its own thread in a shared workspace, where a session of Claude (an instance) can read the program's email and data, run code, prepare pages and save drafts in the mailbox of one of us." And then: "Claude sent no email".

Over the week, "Claude drafted 17 emails, each in answer to a request from one of us, and we sent 12 of them, each after editing." With its drafts, Claude listed 41 points "for which it had found no written statement of ours, so that we could confirm, change or strike each before sending." When a team member challenged Claude's plan to measure binding for all 113 constructs, Claude defended it within 27 seconds. The team member "overruled this recommendation", and four minutes later Claude had revised the memo, the run order and three unsent drafts from 113 constructs to 89. "The judgments remained with us," the report concludes.

The data section describes a check the other reports do not: "Printed values were traced to these tables, and all values of earlier drafts and some of the new values were recomputed by instances that had not produced them." The timings come from message timestamps, and the report states their limits: "five of the nine full answers could be timed to the minute at best."

The AI-use note at the end is unusually candid:

Under our direction, Claude searched the literature, performed the re-analyses, prepared the figures and drafted the manuscript (Methods). The section on working with Claude was compiled by Claude from its own threads and notes and therefore carries a risk of bias, which we have tried to limit by reporting its errors and the limits of the counts.

Two details matter for reading it. The report names the model only as "Claude", with no version. And its competing-interests statement says that Anthropic acquired Mithril in July 2026, owns the antibody program, sells Claude and employs the corresponding author.

BigHat: the announcement's summary only

Anthropic links BigHat's report, titled "Reasoning in the loop: mechanism and measurement in antibody engineering". When we checked on 27 October, the page returned no text, so we have only Anthropic's summary. Claude, "connected to BigHat's lab data system, Reccy, proposed changes to each molecule." Partway through, the team suspected the standard stability and clumping metrics were misleading. "Claude and the researcher reanalyzed the raw data, then proposed different metrics and a protocol change to test them. BigHat's scientists refined and ran it, and the main predictions held."

What did the human checks catch, and what did they miss?#

Two of the reports measure how well people caught Claude's mistakes, and how well Claude caught theirs.

Mithril examined six of Claude's errors in detail. "Claude found five of them, one after a question from a CRO and two before the affected email was sent, while at least three reached a CRO in emails that we had edited and sent." One was a request for less material per tube than the assay CRO had asked for; it was corrected within a day, before any tube was filled. "Two others, both statements in Claude's review of the production data, went unnoticed for more than 4 days, and one remains unsettled."

The report draws the lesson itself: "These cases suggest that our editing tests the decisions in an email more reliably than the details and premises that it carries."

That sentence describes a problem far beyond drug discovery. A person reading an AI draft checks whether it says the right thing. They are much less likely to recheck every number and assumption underneath it. "Reviewed by a human" is true of every draft Mithril sent, and the report shows what that phrase does and does not cover.

At Enveda the comparison ran the other way. On four batch records, the human reviewers raised 11 distinct actionable issues. CDP raised every one of them in at least one of its runs, and single runs added seven that no human reviewer had raised. About four in five comments, from people and from CDP alike, needed no action. CDP's processing plus a quality check with its comments in hand took 166 minutes in total, against 1,553 minutes for unaided review. And according to Anthropic's post, Enveda's standard human review found one minor error in a regulatory document Claude had drafted.

At BigHat, by Anthropic's account, the check went both ways: Claude and an Anthropic researcher questioned the lab's standard metrics, and BigHat's scientists tested the new ones before relying on them.

Four readable reports, four formats#

Side by side, for the four reports we could read:

None of these is wrong. Each fits what its authors needed. Generate needed to know which design ideas worked. Enveda needs to reconstruct any review for an inspector. Mithril wanted to show a new way of running a drug program, and to be honest about where its account came from.

But side by side, three things stand out.

The records live inside each lab's own systems. Generate's agent runs, Enveda's log and Mithril's threads are all internal. A reader sees the report's summary of them, not the record itself.

Part of each record was put together afterwards. Generate recovered its prompts from archived runs. Mithril's account of the work with Claude was compiled by Claude from its own notes. Both reports say so, which is to their credit. But a record assembled after the fact can only be as good as what happened to be kept.

There are no common fields. One report names models per prompt, another names none. One counts phases, another counts emails. A journal editor, a regulator or another lab cannot line these up without reading each report closely, which is what we just did.

Why the format matters#

Journals already ask for this information. Our comparison of AI disclosure policies found the same core asks across publishers: the tool and its version, what it did, how the output was checked and who takes responsibility. Every policy we compared that addresses authorship says an AI tool cannot be an author. Journal AI policies lists what each publisher asks for.

Regulators are starting to ask too. Enveda's report cites draft EU GMP Annex 22, which "for non-critical use, asks that qualified people remain responsible for each output".

And drug development is slow. Mithril's earliest human trials would start in late 2028 at best, and an approved drug would come years after that. Questions about who proposed a design, which model was used and who checked a protocol change can come up long after the people involved have moved on. A record that answers them should not depend on one lab's archive or on an AI's memory of its own work.

What a shared record could add#

This is the part we work on at Papex. Papex is a record for research done with AI: hypotheses, findings and publications, linked to the people, tools and versions behind them, signed and dated.

For each signed version, a Papex record keeps a list of who did what:

• AI work, as the app reported it, with its date. A line reads, for example, "Reported by Claude on 22 Oct", with "edited by the signer" when the person changed it.
• Work the person states themselves, marked "Declared by the signer", with evidence such as a file or a commit.
• A signature from the person accountable, on the exact version. An AI app connected to Papex can prepare a draft and propose changes. It cannot sign the record or publish it, much as Claude at Mithril could draft emails but not send them.
• An AI-use statement generated from the signed version, worded for a journal's AI-use section. It ends with a fixed line: "Names, models and roles are self-reported. This is the signer's statement, recorded with Papex; it is not a review of the science."

A connected AI app can also leave a dated summary at the end of each working session, private to the owner, so that the next session starts from what was actually done rather than from memory.

What Papex does not fix: an AI's report of its own work is still a self-report. Mithril's caveat applies to Papex records too. What changes is when and how the report is kept. It is dated when the work happens rather than reconstructed later, tied to an exact version, signed by a named person, and published with the record when the record is published, rather than kept inside one lab's systems. Papex does not check the science, and it does not check what the AI says it did.

Open questions#

The reports raise questions that no single lab can settle:

• Should an AI's account of its own work be published as written, edited by the people responsible, or checked against logs first? Mithril published Claude's account with a warning; Enveda keeps the logs themselves.
• When several models are given the same prompt, as in Generate's first round, what should the record say about where a design came from?
• How long should agent logs be kept, and who may read them? Enveda keeps them in a regulated system. Most research labs have no such rule.
• When a person edits an AI draft, what exactly have they checked? Mithril's answer, the decisions more reliably than the details, deserves to be tested elsewhere.
• Should reports name the model version, as Generate, Divergence and Enveda do, or is "Claude" enough?

Anthropic says it plans to scale this work up substantially. As AI does more of the work in drug programs, these questions will matter more. The reports published today are a good start: four labs chose to say what Claude did, and one said plainly that its own account might be biased. The next step is a record of who did what that readers can check, not only a description in a report.

If you use an AI app in your research, our guide to AI apps shows how to connect one to Papex and keep that record as you work.

Sources#

• Early progress on accelerating drug development with Claude (Anthropic, 27 October 2026), checked 27 October 2026
• Bai, Kumar, Desai, Li, de Alwis, Hazuda, Grigoryan, Wang and Yi, Refactored intracellular domains enhance CAR expression and function (Generate Biomedicines), checked 27 October 2026
• Erives, Ferrara, Michael, Mallipeddi and Jing, Towards a best-in-class IgE disruptor for food allergy with Claude-enabled antibody design (Divergence Labs, preprint), checked 27 October 2026
• Martin, Schenck, Wee and Colluru, CDP: a language-model system for quality review of executed pharmaceutical batch records (Enveda, 27 October 2026), checked 27 October 2026
• Tranbarger and Mithril Medicine, Preclinical Characterization of a Bispecific Bridging Antibody for the Treatment of LAMA2-Related Congenital Muscular Dystrophy (Anthropic technical report, 26 October 2026), checked 27 October 2026
• BigHat Biosciences, Reasoning in the loop: mechanism and measurement in antibody engineering, linked from Anthropic's post; the page returned no text when checked on 27 October 2026
• enveda/open-doc-review (GitHub), linked from Enveda's report, checked 27 October 2026