Anthropic Is Now Running Its Own Drug Discovery Program Alongside Six Drug Developers

Cody Tranbarger ran a drug company on his own. Mithril Medicine had one employee, two rare forms of muscular dystrophy in its sights and experiments running at 11 outside labs. Claude managed the contracts and logistics and took on much of the company's day-to-day work. In under 18 months, one of Mithril's candidates went from an idea to promising results in mice.

Anthropic now owns Mithril. The company disclosed the acquisition Tuesday in a research update on its drug work, and said Tranbarger will keep running the program from inside Anthropic. If his candidates keep holding up, Anthropic expects one could enter human trials as soon as late 2028.

The same post covers several months spent inside drug programs at six outside developers: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. It arrived a week after OpenAI bought Orchestra Bio to manage the operational side of drug programs. Several of Anthropic's results come from that same side.

The Lab Results

Anthropic's case is that drug development has no single bottleneck, so it put Claude into programs at different stages. Three partners worked on the molecules themselves.

At Generate Biomedicines, six scientists used Claude Code to redesign part of the genetic instructions in CAR T-cell therapy. The treatment has mostly worked against blood cancers, and Generate wants to free up room in those instructions for additions aimed at solid tumors. One Claude design was about 70% shorter than the standard version. In lab tests, immune cells carrying it made roughly 10 times as many cancer-spotting receptors, killed tumor cells more efficiently and stayed active longer.

Divergence Labs is designing an antibody for food allergies that would strip the allergy molecule IgE off immune cells, where current drugs mostly neutralize what floats in the blood. The design needs two working parts. Claude wrote the design plan and then, largely on its own, produced about 35,000 candidate antibodies in roughly a month. Early tests found three that bind IgE and 20 that dislodge it, three of which outperformed the standard medication. Divergence hasn't yet combined the two halves into one drug.

BigHat Biosciences gave Claude access to its lab data system and asked it to improve two antibody fragments. BigHat's automated lab built and tested about 1,010 variants in 10 weeks. Partway through, Claude and an Anthropic researcher concluded that the standard tests for heat stability and clumping were giving misleading readings, and they proposed new measurements. BigHat ran them, and the predictions held. The best variants bound more than 10 times more tightly than the originals.

The Paperwork Results

The other half of the update is about work that never touches a molecule.

Enveda's manufacturing chemists were spending up to 75% of their time on documentation. The company connected its batch records to Claude and had it flag problems. Across 1,061 pages, Claude caught all 11 issues human reviewers had found, plus seven they missed, and cut review time from more than 25 hours to under three. Claude also drafted two regulatory filings. One took four hours. By hand, it would have taken about three months.

Bristol Myers Squibb, the largest name on the list, is using Claude to review trial plans for an early-stage drug and estimate the odds of success. Michael Pourdehnad, who leads early clinical development in hematology, oncology and cell therapy, said that "from a single prompt, Claude generated a robust initial analysis" that would otherwise have taken days. The companies plan to expand the work in 2027.

The Column Group, a biotech investor, is testing whether Claude can shrink the cost of running a company. Its new startup, Pleiades, plans to run more drug programs than a small team usually could, with Claude carrying much of the scientific and administrative load. That project is the least developed of the six. Its first Claude tool, which judges whether a drug target is worth pursuing, has been checked against 45 target-disease pairs.

The Gap Between the Labs Is Closing

When we covered the Orchestra deal last week, the two labs had made different bets. Anthropic was concentrating on early research and doing some of the science itself. OpenAI was buying the operational software that runs inside drug programs.

Tuesday's update shows Anthropic working both ends. Mithril's 11-lab coordination, Enveda's batch records, Pleiades' lean staffing model and BMS's trial planning are all operations work. That is the same part of the timeline OpenAI paid for with Orchestra.

The labs are also selling to the same buyers. Novo Nordisk signed with OpenAI in April and with Anthropic in September.

Owning Mithril raises a question pharma customers have already asked. Reuters reported in September that some of Anthropic's drug industry customers worried the company could learn from their competing programs, even with their data walled off. Anthropic said in June that its own drug work would focus on neglected diseases that big drugmakers tend to skip. Muscular dystrophy fits that description. Whether every future program does is something customers will be watching.

What the Numbers Don't Prove

Anthropic calls the work early, and it is. Almost all of it is preclinical. Divergence's drug hasn't been assembled, and Mithril's best evidence comes from mice.

The figures are also self-reported. Partners co-wrote technical reports, but none of the results have gone through independent peer review. The post says little about where Claude fell short, beyond noting that Anthropic has already prioritized model fixes based on what it learned.

The post also leaves out what Anthropic will do with a drug that works. It doesn't say whether the company would take a candidate to market or license it to a drugmaker.

Where This Goes

Anthropic says it plans to scale the work up substantially. For drugmakers, that means buying from an AI supplier that also runs a pipeline of its own.

The four-hour regulatory filing will get the attention, because time saved is easy to count. The harder test is the one every AI drug effort still faces. No fully AI-discovered drug has cleared a Phase III trial. Mithril's candidate is Anthropic's first chance to change that, and it won't reach people before late 2028.