Anthropic launched Claude Science in beta on June 30, alongside an internal drug discovery program focused specifically on neglected diseases: conditions where the underlying biology is often well understood, but the economics of traditional drug development are unattractive to large pharmaceutical companies (Claude Science, an AI workbench for scientists — Anthropic).
What Claude Science actually is
It’s a research environment built on top of Anthropic’s Claude models, structured around a central coordinating agent that spawns specialized sub-agents for specific research tasks. It ships with over 60 curated skills and connectors pre-configured for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics, and it integrates directly with existing life-sciences models including Evo 2, Boltz-2, and OpenFold3 rather than trying to replace them. A separate reviewer agent inspects every output, flagging incorrect citations, untraceable numbers, and figures that don’t match their underlying code — a specific answer to the obvious worry with agentic science tools, that a model quietly fabricates a plausible-looking result. Every output also carries an auditable history of how it was produced, and the platform manages its own compute, scaling an analysis from a laptop up to hundreds of GPUs on demand. It shipped in beta for Claude Pro, Max, Team, and Enterprise users on macOS and Linux.
Named early users give a clearer sense of what this looks like in practice than the feature list does: Manifold Bio used it for target nomination in tissue-targeting medicine design, the Allen Institute (per researcher Jérôme Lecoq) built multi-agent templates for computational review writing, and UCSF’s Brain Tumor Center (per Stephen Francis) applied it to molecular epidemiology research. Anthropic is also offering roughly $30,000 in compute credits to about 50 selected research projects, with applications closing July 15 and recipients announced July 31.
This is a concrete instance of the generative AI application we’ve written about more generally in our explainer on generative AI in industry: using a model to compress the early candidate-generation and analysis phase of a long, expensive research pipeline, not to replace the wet-lab work, trials, or regulatory process that still has to happen after.
Why the internal drug program is the more interesting part
Running Claude Science as a product is one announcement. Anthropic simultaneously committing to its own internal, preclinical-stage drug discovery work aimed at neglected diseases is a different kind of move. Eric Kauderer-Abrams, Anthropic’s head of life sciences, has framed the two announcements as inseparable: the company says it needs direct, hands-on drug development experience to build a genuinely useful tool for the industry, rather than shipping a platform and hoping customers find the killer use case unassisted (CNBC). Anthropic’s structure as a public benefit corporation is doing real work here too — it gives the company latitude to fund therapeutic programs that don’t have a clear commercial return, which a conventionally structured pharma company would have a much harder time justifying to shareholders.
The competitive landscape
Anthropic is a late but not lonely entrant here. Google DeepMind’s spinoff Isomorphic Labs has been building on the AlphaFold lineage for years and now runs its own IsoDDE drug-design engine, with 17 active drug programs across oncology, immunology, and cardiovascular disease, and the first AI-designed cancer drug expected to enter Phase 1 trials by the end of 2026 (Isomorphic Labs). OpenAI has its own life-sciences model, GPT-Rosalind, currently in research preview through a “trusted access program” for vetted biomedical research organizations (OpenAI).
| Lab | Product | Current stage | Distinct focus |
|---|---|---|---|
| Anthropic | Claude Science + internal drug program | Public beta (Jun 30, 2026) | General-purpose research workbench across 60+ tools; own neglected-disease pipeline |
| Google DeepMind / Isomorphic Labs | IsoDDE | 17 active drug programs; first candidate entering Phase 1 by end of 2026 | Structure-prediction-first, built on the AlphaFold lineage |
| OpenAI | GPT-Rosalind | Research preview, invite-only | Reasoning model tuned specifically for biology/chemistry/genomics workflows |
Worth watching over the next year: whether “neglected diseases” research stays a genuine focus once the commercial product side of Claude Science starts generating meaningful revenue, or whether it quietly narrows toward the higher-margin pharma R&D use cases that funded it in the first place. That tension, between the framing used to launch a product and where the commercial incentives actually point afterward, is close to a universal pattern in big-lab AI announcements with a social-good angle.
Why drug discovery is a harder target than it looks
It’s worth being honest about why this category is difficult even for a very capable model. The traditional drug discovery timeline is widely cited across the industry as running a decade or more and costing well over a billion dollars per approved drug, and the reason isn’t primarily a shortage of candidate molecules. It’s that almost every candidate fails somewhere in preclinical toxicity testing or clinical trials, for reasons that often aren’t predictable from molecular structure alone. An AI system that gets meaningfully better at proposing and prioritizing candidates compresses the front end of that pipeline, but it doesn’t touch the part that actually consumes most of the time and money: proving a candidate is safe and effective in living systems, which still requires the trials themselves.
That’s the realistic frame for judging Claude Science’s eventual impact. If it measurably shortens the candidate-generation phase for neglected-disease research, that’s a genuine and valuable outcome, especially for diseases too commercially unattractive to get this kind of tooling investment otherwise. But the honest way to evaluate the announcement over the next year isn’t “did Claude Science discover a drug,” it’s whether Manifold Bio, the Allen Institute, UCSF, or Anthropic’s own internal program can point to a candidate that made it further into the pipeline, faster, than it would have without the platform — and whether Isomorphic Labs’ Phase 1 candidate reaches the clinic on schedule, which would be the first real proof point for this entire category of tool, from any lab. That’s a much higher and more specific bar than a product launch keynote, and it’s the one worth holding this to.
