Claude's Six-Month Drug R&D Report Card: Designing Molecules, Managing Experiments, Writing Filings—Earliest Human Trials by Late 2028

On October 27, Anthropic published a post describing what Claude has done in drug discovery over the past few months. Here is the conclusion up front: these projects are basically still at the laboratory and animal-testing stage, and none of the drugs has been tested in humans. For the fastest-moving candidate, human clinical trials may begin no earlier than late 2028.

A new drug usually takes more than a decade to go from idea to market: identifying targets, designing molecules, iterating and optimizing, solving manufacturing at scale, writing regulatory filings, and running clinical trials. The often-heard 'AI designs molecules' is only one part of this; Anthropic wants Claude to participate in every step.

What Each Company Had Claude Do

The cancer therapy CAR-T works by installing a 'genetic instruction manual' into a patient's immune cells so they can recognize cancer cells. Scientists at Generate Biomedicines used Claude Code to shorten one section of the manual by about 70%, freeing up space to add other instructions so that it could potentially also treat solid tumors in the future. In the lab, the new version actually performed better: the number of 'probes' on the immune cell surface that recognize cancer cells increased by about 10x.

Divergence Labs is working on a food allergy drug and needs a molecule that can pry allergy-triggering IgE off immune cells. Claude essentially did this on its own, generating about 35,000 designs in one month; among the filtered candidates, 3 could pry off more than existing drugs.

BigHat Biosciences had Claude suggest modifications for two drug candidates, and in 10 weeks it made and tested about 1,010 variants. The best version bound its target more than 10 times more tightly, and one of them clumped more than 100 times more slowly, making it easier to manufacture and store.

Halfway through, the team suspected that standard thermal stability and clumping metrics were misleading; Claude then went back and recalculated the raw data, switched to a different set of metrics, and the experimental results validated the prediction. This is even more hardcore than generating tens of thousands of designs.

At Enveda, production chemists spend up to 75% of their time on paperwork. Claude reviewed a 1,061-page production document, found all 11 issues that humans had found, and additionally found 7 issues humans had missed, cutting review time from more than 25 hours to under 3 hours. For an important regulatory filing, Claude produced a first draft in 4 hours; doing it manually takes about 3 months.

In addition, Bristol Myers Squibb (BMS) is using Claude to evaluate clinical trial protocols for early-stage drugs, and investment firm The Column Group is establishing a new company run with Claude that will let a small team advance many drug programs at once.

A One-Person Drug Company

Anthropic also acquired Mithril Medicine, a drug company with just one person. Founder Cody Tranbarger studies rare childhood muscular dystrophy; contracts, logistics, and experiment scheduling with 11 external labs were all handed over to Claude, and in less than 18 months he took a therapeutic idea for one disease from concept to showing efficacy in mice.

This disease is called LAMA2 congenital muscular dystrophy. Affected children are born lacking a protein that anchors muscle fibers, and their muscles begin to break down from birth. Mithril designed a two-headed antibody that re-'pins' another similar protein remaining in muscle to the muscle. In mouse experiments, the treated group's forelimb strength and grip strength were about twice those of the untreated group.

How Much Should We Trust These Results

The results all come from technical reports and preprints released by the partners and by Anthropic itself; the article does not mention any independent external verification. Mithril's report is marked as not peer-reviewed, and the mouse experiment is preliminary: 7 mice in the treated group and 4 in the control group. After comparing by litter, the difference was still slightly short of the usual statistical significance threshold, and the report itself says rigorous replication experiments are needed to confirm it.

The report also describes Claude's mistakes: the authors examined 6 errors in a week of Claude's work. Five were found by Claude itself, but at least 3 had already been sent out in emails that humans had reviewed and edited, and two of those were discovered only after more than 4 days. Claude only drafts emails; humans decide whether to send them.

If you focus only on 'AI designing molecules,' you'll miss the more interesting part. In this collaboration, the most obvious efficiency gains are in paperwork and operations: reviewing documents, writing filings, and coordinating labs. These steps are not glamorous, but they really do save time. We are still far from an 'AI drug god,' but a tool that can handle coordination across 11 labs and compress a three-month regulatory draft into 4 hours has already begun changing how some people work.

Source: Anthropic Blog