Anthropic is entering science. Clinical research is on the map.

On June 30, Anthropic launched Claude Science, a workbench built to run research rather than assist with it: literature analysis, coded analysis pipelines, figures, and manuscript drafts in one environment, with the model executing multi-step workflows across more than 60 scientific databases. The same week, the company said it is starting its own preclinical drug discovery program, aimed at targets it says traditional pharma finds unattractive. A month earlier it hired John Jumper, who shared a Nobel Prize for AlphaFold, away from Google DeepMind.

None of that is aimed at a monitoring visit. All of it is aimed at the industry that pays for monitoring visits, which is why it belongs in this trade daily.

What actually launched

Claude Science is in beta on macOS and Linux for paid Claude plans. The pitch is workflow, not a new model: connectors to UniProt, PDB, Ensembl, ChEMBL, and similar databases, prebuilt skills for genomics, proteomics, and cheminformatics, compute that scales from a laptop to rented GPU clusters, and a reviewer agent that checks citations and calculations in the output. Every result ships with its code and environment so someone else can rerun it. Anthropic is also funding up to 50 academic projects with credits worth up to $30,000 each.

That is bench science and dry-lab analysis. The clinical layer sits in the company's earlier releases: since January, Claude's life sciences connectors have included Medidata for enrollment and site performance data and ClinicalTrials.gov for pipelines and site selection, with protocol drafting and regulatory submission preparation named as intended uses. Novo Nordisk, Sanofi, and Genmab appear on the customer list. Bristol Myers Squibb and Genentech demoed at the launch event.

The one deployment with numbers

Most AI-in-pharma claims arrive without a baseline. One does not. Novo Nordisk has said publicly that its NovoScribe system, built on Claude, cut clinical study report drafting from roughly 12 weeks to about 10 minutes, and that the writing team on that work went from more than 50 people to 3, with humans moved to review and approval. Those are company-reported figures, not an independent audit, and drafting is not the same as a finished, signed CSR. But a headcount that goes from 50 to 3 is not a marketing percentage. It is an org chart.

Medical writing is the first clinical research function where the before and after are both on the record. Our salary survey counts medical writing as its own function for exactly this reason.

Where the moat still holds

A workbench that can draft a protocol is not a validated system. Sponsor-side deployment runs through GxP: audit trails, Part 11 compliance, documented validation, and a named human who signs. Nothing announced changes that. The pattern from statistical programming, which we covered last week, repeats here: the documentation-shaped work moves first, the verification and the signature stay human, and the job shifts toward specifying work precisely and owning the QC that proves it right.

It is also worth reading the direction of travel honestly. Anthropic's connector list touches feasibility, enrollment tracking, site selection, protocol drafting, and submission prep. It does not touch consenting a patient, managing a site relationship, or finding the discrepancy a system swears is not there. Vendors go where the text is. The parts of clinical research that are mostly documents should expect company, and soon.

What to watch

Three things will tell you whether this is a wave or a press cycle. First, whether other sponsors publish Novo-style numbers with baselines attached. Second, whether CRO job postings for medical writing and regulatory roles shift toward review and oversight language. Third, what happens to headcount, which shows up in WARN filings and in this community before it shows up in earnings calls.

No CRO or sponsor has announced clinical research layoffs attributed to AI, and we could not verify any. What the tools pay, and what they change, will show up first in self-reported datapoints. If you write, program, or file for a living, your datapoint is the measurement.