Over several months, Anthropic's scientists worked with six drug companies and investors, including Bristol Myers Squibb and Enveda, and bought a one-person muscular dystrophy startup to run drug programs of its own. The full set of results is mostly about molecule design, but the part most readers will recognize is the work around the molecules.
Reviewing records is where the hours go
Making a drug under regulation means keeping a record of everything done to every batch, and a quality team reads those records before the batch can be released. At Enveda, a manufacturing chemist can spend about 75% of their time on this reading, and one senior reviewer needed 10 hours for a single 500-page record.
Enveda connected Claude to those records. Claude works out which batch and which approved instructions a document belongs to, checks each page against them, and flags errors and open questions. It sees a read-only copy and cannot change the official record, and Enveda's reviewers decide what to do with every flag. On four records of 1,061 pages in all, the system raised all 11 issues the human reviewers had found. In its timed run it also raised seven more, which Enveda's quality team took up with the manufacturer before the batch was released. Review time fell by 89%, from about 26 hours to under three.
Enveda also had Claude draft the summary section of an application to the FDA. The first draft took four hours, against about three months by hand, and people still review it afterwards. The useful design choice was to have Claude first pull the key facts into one shared dataset and then write every document from that. When a number changes, it changes everywhere at once.
One person, eleven labs
Cody Tranbarger ran Mithril Medicine alone, with outside consultants and 11 outside labs. When responsibility for a set of experiments changed hands, Claude pieced together where things stood from email threads and 14 months of slide decks, and drafted five handover emails that he reviewed and sent. When the labs' screening results came back as hundreds of files, Claude combined them into one list of candidates. Mithril took a possible treatment from idea to promising results in mice in under 18 months, and turning a new idea into a full design plan took two to three days, where he estimates one to two months without Claude. So far these are mouse results, not a medicine.
The bigger consequence is for diseases with few patients. One of Mithril's targets, a muscle disease called LAMA2-related congenital muscular dystrophy, affects an estimated 4,000 to 5,000 people in the US and Europe. A full-sized team is hard to justify for that few patients, and most of what such a team does is coordination. If one scientist and Claude can carry that load, the smallest market that can pay for a drug program gets smaller, and some diseases that were never worth attempting become worth it.
The judgment stays with people
The limits are visible too. Divergence Labs, working on a food allergy drug, had Claude generate about 35,000 antibody designs and got lab results in weeks, where comparable work had taken up to nine months. But after one question had a clear answer, Claude proposed five more design runs that were unlikely to change it, while the human researcher suggested a different strategy. At BigHat Biosciences, a researcher suspected the lab's stability scores were wrong, and Claude then went through the raw instrument readings and worked out why. The people supplied the suspicion and chose what was worth testing. Claude did the reading.
What this looks like outside a lab
The spreadsheet case ran the same way. Since 1980 there are about 400,000 fewer accounting clerks and about 600,000 more accountants. The clerk's work was the checkable part, the part with a right answer on the page, and once that was cheap, demand for the judgment on top of it grew. In drug development, cheaper review should mean more programs per reviewer, not fewer reviewers.
Take a hospital finance team that spends 20 hours a month checking supplier invoices against contracts. The same pattern would work: a fixed rule finds the contract that governs each invoice, Claude flags prices and terms that do not match, it works on copies and cannot change the books, and a person decides what to dispute. If that saved anything close to Enveda's 89%, the team would get back about 18 hours a month, more than two working days, for the negotiations the flags point to.
The place to start is the check you already do every month against a written standard. That is the work to hand over first, and the decision at the end of it is the part to keep.