Claude caught every problem human reviewers had found in 1,061 pages of drug-manufacturing records, and seven more they had missed, according to results Anthropic published on 27 October. The figure comes from Enveda, one of six organisations that have spent the past several months using the AI model on real drug programmes, from designing molecules to planning clinical trials.
In a research update, Anthropic named the partners as BigHat Biosciences, Bristol Myers Squibb (BMS), Divergence Labs, Enveda, Generate Biomedicines and The Column Group. It also disclosed that it has acquired Mithril Medicine, a one-person company working on two rare forms of muscular dystrophy. No price or terms were given. The results are reported by the companies themselves, and almost all of them are from the lab rather than the clinic. As Anthropic puts it: “All of this research is still at an early stage.”
The update builds on smaller experiments the company has published this month, including the campaign in which Claude designed protein binders against scorpion neurotoxins. This time the work sits inside other companies’ pipelines, and the claimed gains are as much about paperwork and lab cycles as about new molecules.
In brief
- What happened: Anthropic published early results from Claude work with six drug developers and said it has acquired Mithril Medicine, for an undisclosed sum.
- Why it matters: the claimed gains span the whole pipeline, from antibody and CAR T design to manufacturing records, where Enveda says review time fell 89%.
- What’s unknown: the results are company-reported and mostly preclinical; no drug is in human trials, and Anthropic’s late-2028 trial date is an expectation.
Six programmes, six different bottlenecks
Anthropic argues that “the bottlenecks are distributed all throughout the drug development process”, so it chose partners working at different stages. The headline results, as each company reports them, are below.
| Partner | Stage | Reported result |
|---|---|---|
| Generate Biomedicines | CAR T design | Signalling part about 70% shorter; more receptors, better tumour-cell killing in lab tests |
| Divergence Labs | Food-allergy antibodies | About 35,000 designs in about a month; 3 hits binding IgE, 20 dislodging it |
| BigHat Biosciences | Antibody optimisation | Tighter binding, stable at up to 18°C higher, much slower clumping |
| Enveda | Manufacturing records and filings | Human review time cut 89%, from over 25 hours to less than three |
| Mithril Medicine | Muscular dystrophy | Idea to promising results in mice in less than 18 months |
Generate is trying to adapt CAR T-cell therapy, which has mostly helped blood cancers, so it can work against solid tumours. Its scientists used Claude Code to redesign the receptor’s internal signalling section. In its own blog post, Generate says its best design is “just 45 amino acids, 70% shorter than the second-generation design” and that it chose 186 designs for the first round of testing. Anthropic says the design let immune cells make “roughly 10 times as many receptors”, killed tumour cells more efficiently and stayed active for longer, all in lab tests.
Divergence Labs wants to pull the allergy molecule IgE off immune cells altogether, rather than only mopping up the IgE in the blood as standard medicines do. Anthropic says Claude, “largely autonomously, generated roughly 35,000 antibody designs in about a month”. Early tests found three designs that bound IgE and 20 that dislodged it, three of which “appeared to dislodge IgE more than the standard medication”. The next step is to combine the two halves into one drug.
BigHat’s own figures differ from Anthropic’s
At BigHat, Claude was connected to the company’s automated lab through its data system, Reccy, to improve two antibody fragments. Anthropic says the lab made and tested “about 1,010 variants in about 10 weeks” and that the best versions bound “more than 10x more tightly” than the originals. BigHat’s own account gives different numbers: 1,232 antibodies tested over seven design rounds, with the best variants binding “up to nine times more tightly”. Neither post explains the gap. Both agree on the 18°C gain in heat stability; Anthropic says the other lead clumped “more than 100x more slowly”.
BigHat adds a detail that matters for anyone wondering how much the model did on its own. Claude and an Anthropic researcher noticed that a standard instrument was making misleading calls on stability, proposed a different heating protocol, and BigHat’s scientists ran it. The company says each design round took about a week, against “at least 3-5x longer” for similar lab cycles.
Enveda’s results are the least glamorous and the most concrete. Its manufacturing chemists, Anthropic says, spent “up to 75% of their time on paperwork”. Claude also drafted two regulatory documents; one took four hours, against “about three months by hand”, and Enveda’s normal review of the other found one minor error.
The remaining two partners are earlier on. The Column Group, an investor, is setting up a company called Pleiades to run more drug programmes with a small team; its first Claude skill, for judging drug targets, has been tested on 45 target-disease pairs. BMS is using Claude to assess trial plans for an early-stage drug and estimate its chances of success, with plans to expand the work in 2027. Michael Pourdehnad, who heads early clinical development in haematology, oncology and cell therapy at BMS, said: “From a single prompt, Claude generated a robust initial analysis that helped inform our assessment of different development options, which would otherwise take days to complete.”
A one-person biotech and a 2028 target
Mithril Medicine is the most striking piece of the announcement. Its founder, Cody Tranbarger, used Claude to organise contracts, logistics and experiments with 11 external labs, according to Anthropic, and took a possible treatment for one of the two conditions from an idea to “promising results in mice in less than 18 months”. He is now at Anthropic and plans to start several new programmes. The company’s expectation is conditional: “If his drug candidates continue to hold up, we expect a drug to be ready to start clinical trials in humans as soon as late 2028.”
Anthropic is not the only AI lab buying its way into drug development. On 20 October OpenAI said it had acquired Orchestra Bio, whose software coordinates drug programmes, to fold into its Rosalind life-sciences suite.
The clearest wins are in the paperwork
The temptation with an update like this is to read it as AI discovering drugs. The evidence points somewhere more modest and, in the near term, possibly more useful. The most measurable gains are in work that rarely makes headlines: checking batch records, drafting filings, turning around lab cycles faster. An 89% cut in review time on manufacturing documents, if it holds up elsewhere, saves weeks on the route to a trial without anyone needing to trust a molecule the model dreamed up.
The design results are harder to judge. Every one is a lab or animal result chosen and reported by the companies involved, and the BigHat discrepancy shows why the underlying reports matter more than the summaries. Independent studies such as the AbOptimizer paper on AI-edited antibodies and Pythia Labs’ preprint on antibodies designed from scratch suggest that tighter binding in a test tube is becoming routine. Turning that into a medicine that is safe and works in people is still the hard part, and nothing here has reached that stage.
The Mithril story raises a different question. If one founder with a model can coordinate 11 outside labs, the cost of starting a drug programme could fall sharply, which is the bet behind Pleiades too. Whether that leads to more cures, or simply more early-stage candidates competing for the same trial funding, will not be clear for years. None of the announcements mentions UK sites or regulators.
Clinical work is the next test
Anthropic says it will extend Claude into clinical trials as these programmes advance and as it takes on partners whose drugs are already in human testing. It has also launched Claude Science, a workbench for researchers, along with Claude credits for life scientists, a Life Sciences Verification Program that gives vetted organisations looser safeguards, and an open call for guest investigators. Generate says it expects experimental results on solid-tumour targets chosen with Claude “within the next six months”. The firmest date on the table is still Anthropic’s own: late 2028, at the earliest, for Mithril’s first human trial.
Sources
- Anthropic: Early progress on accelerating drug development with Claude (27 October 2026)
- BigHat Biosciences: Agents at the bench: Anthropic and BigHat collaborate on rapid agentic-driven therapeutic antibody optimization (27 October 2026)
- Generate Biomedicines: Less is More: Refactoring the CAR Signaling Domain (27 October 2026)
Related stories
- Anthropic Says Claude Designed Protein Binders Against Scorpion Neurotoxins
- OpenAI Buys Orchestra Bio to Put Drug-Programme Agents in Rosalind
- Pythia Labs Launches With AI Antibodies That Bound 15 of 22 Targets
- AI-Edited Antibody Bound IL-6 13 Times More Tightly in Lab Tests
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
