Claude cut a drug company's document check from 25 hours to under three

The paperwork result

As TechRadar reported, Anthropic is pushing Claude deeper into drug development, and the most concrete figure comes from a company called Enveda. Documentation for making drugs had eaten up to 75% of some chemists' working time there, so Enveda tried Claude on it.

In a test covering 1,061 pages, Claude flagged all 11 problems that human reviewers had already found. It also turned up seven more. Manual checking dropped by 89%, from more than 25 hours to under three.

That is the headline number, and it is a believable kind of win. Nobody is claiming the AI invented a cure. It read a very long stack of documents and caught inconsistencies, a chore that is slow and tedious for people.

Lab results from three biotech partners

Anthropic's blog post, as summarised by TechRadar, goes well beyond paperwork. It describes work on molecule design, where the aim is to find a protein or cell part that does a specific job.

• Generate Biomedicines: a team of six scientists used Claude Code, Anthropic's coding tool, to redesign a piece of a CAR T-cell therapy. CAR T therapy engineers a patient's immune cells to attack cancer and is traditionally used against blood cancers. A shorter part leaves room for extra genetic instructions that might help it work on solid tumours. The new part was about 70% shorter, and the cells made roughly ten times more receptors, the surface structures that spot cancer. In lab tests they also killed cancer cells more effectively and stayed active longer.
• Divergence Labs: Claude produced about 35,000 molecular designs in a month for a possible food-allergy treatment aimed at IgE, an immune molecule. Early tests found three designs that bind IgE and 20 that pull it off immune cells. Three of those 20 looked better than the standard medication at removing it.
• BigHat Biosciences: paired with automated lab equipment, Claude helped improve two antibody fragments. Across about 1,010 variations tested over ten weeks, the best cancer-targeting version gripped its target more than 10 times as tightly and stayed stable at temperatures 18°C higher. The other fragment clumped together more than 100 times more slowly.

Earlier decisions at Bristol Myers Squibb

Anthropic is also working with Bristol Myers Squibb on a different problem: judging plans for clinical studies and estimating how likely they are to succeed. Michael Pourdehnad, the company's senior vice president for early clinical development, said, "From a single prompt, Claude generated a robust initial analysis."

He said this helped the team weigh options that would otherwise have taken several days. The two companies plan to look at whether Claude can help with decisions in early clinical development more broadly.

This is a case where the AI supports a human decision rather than making it. The value is a faster first draft of the analysis, which experts then judge.

How much to trust the numbers

All of these figures come from Anthropic's account and its partners, not from independent testing. TechRadar notes that most of the treatments are still at experimental stages, so the biology results are lab findings, not results in patients.

The 89% figure is also narrow. It measures one review task at one company. A saving that large on a 1,061-page review does not mean every document job shrinks the same way.

Still, the setup is a good template. Enveda had a clear baseline (more than 25 hours), a known list of problems to check against (the 11 humans found), and a measurable outcome. Anthropic says its approach is to get deeply involved in real drug programs, because that is how it finds the bottlenecks and the weak spots in its own models. It plans more partnerships, research programmes and tools for life-science organisations.