Anthropic plugs Claude into drug development: 6 partners and a CAR-T that's 70% shorter

⚡ The gist in 5 seconds

• Anthropic has plugged Claude into real drug development programs at six partners and launched its own.
• One CAR-T design proposed via Claude Code came out 70% shorter than the standard one and performed better in the lab.
• It's all preclinical for now: technical reports, preprints, and a promise to sharply scale up the work.

🔍 What happened

Anthropic's biology and medicine team published a report on several months of work with six organizations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group. In parallel, the company has started its own drug discovery programs — including for rare childhood forms of muscular dystrophy, where Claude is handling most of the scientific and operational work.

The mechanics are visible in the Generate Biomedicines example. The company is reworking a CAR-T therapy that usually targets blood cancers so it also works on solid tumors. To do that, part of the genetic instruction needs to be shortened to free up space for new ones. Six scientists ran variants through Claude Code: one design came out roughly 70% shorter than the standard one. In lab tests, immune cells carrying it produced about 10 times more receptors, killed tumor cells more effectively, and stayed active longer. Divergence Labs has a different task — food allergies. There, they're designing a two-armed molecule that strips IgE off immune cells entirely, rather than just neutralizing it in the blood like standard drugs do.

🧠 Our take

Our assessment: Anthropic has spent years proving its models are useful in code, and medicine is the next obvious market — long payoffs and a high cost of error. Partnerships with pharma deliver what no benchmark can: real tasks, real data, and case studies you can show off. The company itself writes outright that some model improvements have already been prioritized based on these projects.

What this changes. The 70% story isn't about "AI invented a drug" — it's about engineering: shortening a genetic construct without breaking it is painstaking work, and here the model genuinely saved six scientists time. But a shorter instruction and ten times more receptors in cell culture is still not a drug that reaches a patient. Between those two points lie years and money, and that's usually where everything dies.

Who wins. Biotechs with their own data and access to strong models — they get an acceleration in the early stages. Who loses — companies selling an "AI platform for drug discovery" without published case studies: now there's something to compare against. We believe that publishing reports and preprints matters more here than the numbers themselves: it sets a transparency bar for the whole market.

⚠️ What we don't know

The post contains no sample sizes, no cell line names, no timelines, no money. "More effective" in lab tests is a phrase that could mean anything, and the preprints, as usual, haven't been peer-reviewed. About its own muscular dystrophy programs, Anthropic said almost nothing: no stage, no results. And most importantly — "we're optimistic that some programs will lead to working medicines" is a hope, not data. What's worth checking are the preprints, not the blog summary.

Prediction: within two quarters, Anthropic or its partners will publish another report with new preclinical numbers — the company explicitly promises to scale up the work. We don't expect these programs to move into clinical trials within the same horizon.

Source: Anthropic