What Anthropic launched
On June 30, 2026 Anthropic unveiled Claude Science. The idea, in one line: to do for scientific research what Claude Code did for software development.
It's not a chatbot for researchers. It's a working environment — a workbench — where the AI carries out real work from high-level instructions: it analyzes the literature, sets up and runs multi-step analyses, refines figures and manuscripts. And it's available right away, in beta, for all paid plans (Pro, Max, Team, Enterprise), on macOS and Linux.
Anthropic also added a grants program: up to 50 projects funded with credits of up to $30,000, applications open through July 15, projects running from September to December. A way to get the tool into academic labs quickly.
How it actually works
At the center is a coordinating agent, a kind of project manager, with access to over 60 skills and connectors already configured for research domains: genomics, single-cell, proteomics, structural biology, cheminformatics.
This agent is linked to the databases researchers use every day — UniProt, PDB, Ensembl, ClinVar, ChEMBL, GEO — and can lean on specialized biology models like Evo 2, Boltz-2 and OpenFold3 through NVIDIA's BioNeMo toolkit. You can create custom agents for your workflow, and a separate reviewer agent checks citations and calculations.
In practice: instead of jumping between ten disconnected tools, the researcher works in one place and delegates the mechanical steps to the AI, keeping control over the critical ones.
Want to understand where agentic AI can remove friction from your processes?
30 minutes to discuss your specific case.
The real point is reproducibility
This is where it differs from a generic assistant.
Every figure produced by Claude Science carries with it the exact code and environment that generated it, a plain-language description of how it was created, and the full message history. You can modify a result by writing in natural language, and the agent rewrites its own code accordingly.
For anyone working in a regulated context — pharma, biotech, pharmacovigilance — this matters more than any benchmark. A result you can't trace and verify is worthless in those fields. Having it built into the tool, rather than bolted on afterward, is the choice that makes Claude Science genuinely usable in a lab.
The first concrete results
The early numbers come from real labs, and should be taken as indications, not universal promises.
At the Allen Institute, Jérôme Lecoq built a multi-agent computational review template with about twenty custom skills, cutting the writing of a review from roughly two years to a few months. At UCSF's Brain Tumor Center, Stephen Francis accelerated germline variant analysis to about one-tenth of the previous time. Manifold Bio used Claude Science to nominate tissue-targeting drug candidates, assessing surface expression, trafficking and safety.
They're different cases, but the thread is the same: the bottleneck in research is often not the raw power of the model, but the friction of making the AI genuinely useful inside the lab. That's exactly the point Anthropic built the product on.
What it means, even if you don't have a lab
Claude Science is vertical, but the signal is general.
After Claude Code for software and Cowork for office work, Anthropic adds an environment dedicated to an entire sector. The direction is clear: Claude is no longer just a model you ask questions to, but a platform on which to build specialized agents for the work of a specific domain. Anyone running a company should read it this way: the question is no longer "which chatbot do I use," but "which workflows in my sector can I redesign with custom agents."
For those in pharma and pharmacovigilance or healthcare, Claude Science is worth evaluating now. For everyone else, it's yet another confirmation that it's worth asking where agentic AI can remove friction from your own processes — not someday, but with the tools already available today.