Anthropic has published its most detailed account yet of Claude at work in drug development, covering six partner companies over several months and its purchase of a one-person biotech. Most of it is preclinical, and all of it is reported by Anthropic and its partners. It still matters, because it shows where a general-purpose model is already saving time in a slow and costly industry, and where the claims stay unproven.
The partner results span the pipeline. At Generate Biomedicines, six scientists used Claude Code to propose designs for a CAR T-cell therapy, and one design about 70% shorter than the standard version produced roughly 10 times more receptors in the lab. Divergence Labs, working on food allergies, had Claude generate about 35,000 antibody designs in about a month, largely on its own; 3 bound the target and 20 dislodged it from immune cells. At BigHat Biosciences, about 1,010 variants were made and tested in about 10 weeks, with the best binding more than 10 times more tightly. Bristol Myers Squibb, the newest partner, uses Claude to review plans for an early-stage clinical trial.
The paperwork result is the easiest to believe. Enveda, whose manufacturing chemists spent up to 75% of their time on documents, connected its records to Claude. On 1,061 pages, Claude caught all 11 issues human reviewers had found plus 7 they missed, and review time fell 89%, from over 25 hours to under 3. One regulatory draft took 4 hours against about 3 months by hand. Work like this does not depend on a molecule surviving biology, so the gain is real today. Anthropic also bought Mithril Medicine, whose founder used Claude to coordinate 11 outside labs and took a muscular dystrophy candidate from idea to promising mouse results in under 18 months.
The announcement fits a race among the large labs to prove value in science. OpenAI has its GPT-Rosalind life sciences model, and Google DeepMind has added chemistry skills to Antigravity and shown agents proposing an antimicrobial peptide. Anthropic's angle is breadth: rather than a specialist model, it is pushing Claude Code and its general models into every stage from design to filings, and it has launched a verification programme that loosens safeguards for vetted organisations.
The limits are plain. Design counts and lab assays are early signals, and most drug candidates fail in human trials. None of these programmes has reached patients; Mithril's earliest human trials are described as possible from late 2028. The results are selected by the companies involved, with no failures or comparison groups reported, and the partners have commercial reasons to praise the tool. The fair test is how many of these candidates reach the clinic, and how the paperwork savings hold up under regulators' scrutiny.