Claude Science looks like dogfooding for neglected-disease pipelines

TL;DR

Anthropic launched Claude Science on June 30, targeting neglected diseases, and I read it as deliberate dogfooding — the same path Claude Code took from internal tool to product. Claude Science stitches together literature, code, compute, databases, and reviewer agents into a full scientific workbench. The neglected-disease focus lets Anthropic sit inside a pipeline without looking like a competitor to pharma customers. The real bet is that value in AI-bio accrues at the workflow and orchestration layer, not at the model weights.

Anthropic is getting into drug discovery itself. It launched Claude Science on June 30 and said the program targets neglected diseases.

This looks like dogfooding, the way Claude Code lived as an internal tool before it became a product. If Claude is going to be useful in drug discovery, Anthropic needs exposure to the ugly middle: assay design, evidence quality, prioritization, validation, handoffs, and kill decisions. An internal program buys that feedback loop where the evals are slow and expensive.

Claude Science is a workbench for scientists. It stitches together literature, code, compute, scientific databases, figures, manuscripts, and reviewer agents that inspect outputs while a pipeline runs. It connects to NVIDIA's BioNeMo toolkit, Evo 2, Boltz-2, and OpenFold3. It can run on a lab's own laptop, Linux box, or HPC node, so sensitive data does not have to leave the systems it already sits on.

The neglected-disease focus does commercial work too. It lets Anthropic sit inside a pipeline without looking like a competitor to the pharma customers it wants building on Claude. They have not said whether any molecule reaches the clinic, who would own it, or what commercialization looks like. A lot of question marks remain.

The real bet: value in AI-bio accrues at the workflow and orchestration layer more than at the model weights. Thomas Clozel, MD and CEO of OWKIN, drew that line earlier this week. Asked if Owkin can compete, he pointed to the harness: data, wet lab, hospital trust, validation from R&D to care. Model performance was not on his list.

His test: in two years, count the patients treated where nothing was available before. That is exactly the layer Anthropic is now reaching for.

Key takeaways

• I see Claude Science as dogfooding — the same move Anthropic made with Claude Code before it became a product, buying a feedback loop where evals are slow and expensive.
• Claude Science is a full scientific workbench: it stitches together literature, code, compute, scientific databases, figures, manuscripts, and reviewer agents that inspect outputs while a pipeline runs.
• It connects to NVIDIA's BioNeMo toolkit, Evo 2, Boltz-2, and OpenFold3, and can run on a lab's own laptop, Linux box, or HPC node so sensitive data does not have to leave existing systems.
• The neglected-disease focus does commercial work too — it lets Anthropic sit inside a pipeline without looking like a competitor to the pharma customers it wants building on Claude.
• Key questions remain unanswered: whether any molecule reaches the clinic, who would own it, and what commercialization looks like.
• The real bet is that value in AI-bio accrues at the workflow and orchestration layer more than at the model weights — a line Thomas Clozel, MD and CEO of OWKIN, drew independently this week, pointing to data, wet lab, hospital trust, and validation rather than model performance.