Anthropic launches Claude Science to automate lab research

Most biomedical researchers spend less time making discoveries and far more time acting as human data pipelines—jumping across databases, writing conversion scripts, waiting on HPC queues, and hunting down broken citations.

Anthropic has now introduced Claude Science, a dedicated desktop AI workbench for researchers. Available in beta on macOS and Linux for Claude Pro, Max, Team, and Enterprise subscribers, it aims to pull that fragmented scientific workflow into a single command center.

Crucially, Anthropic notes that Claude Science is not a new biological foundation model. It runs the same Claude models available everywhere today, including Claude Opus 4.8. But the wrapper around them is heavily specialized. According to a report by STAT, Anthropic's life sciences head Eric Kauderer-Abrams also revealed at the briefing that the company plans to try developing drugs of its own to stress-test its tools in the real world.

A generalist agent managing an army of specialists

Instead of just a chatbot window, Claude Science behaves like a project manager for the lab bench. A central coordinating agent comes pre-loaded with over 60 skills and connectors spanning genomics, proteomics, single-cell analysis, structural biology, and cheminformatics.

It natively integrates with NVIDIA’s BioNeMo Agent Toolkit. Through plain-language prompts, researchers can invoke domain models like Evo 2, Boltz-2, and OpenFold3, alongside accelerated libraries. NVIDIA notes that tools like RAPIDS-singlecell can compress a 1.3-million-cell clustering pipeline from 52 minutes down to just 25 seconds, while nvMolKit accelerates operations like conformer generation by up to 3,000x.

It also handles compute logistics. Claude Science can run locally, connect via SSH to an on-prem HPC cluster, or dispatch jobs to on-demand cloud platforms like Modal, scaling from a single GPU to hundreds. It drafts an execution plan, requests confirmation before touching remote resources, and executes the code. Because the software runs on lab-controlled infrastructure, proprietary and patient data stays put, and only the context required for a specific step is shared with Claude.

Reproducibility and critic agents

AI hallucinations in academic work have made scientists understandably wary. Anthropic tackles this with an actor-critic pattern: while generating agents produce text and code, an automated reviewer agent inspects the outputs, flagging untraceable statistics, citation hallucinations, and figures that diverge from underlying code.

The app also natively renders interactive 3D protein structures, chemical diagrams, and genome tracks. Every artifact bundles the exact code, execution environment, and prompt history required to reproduce it months later. If a figure needs tweaking, researchers can ask Claude Science to alter axes or remove gridlines in plain English, and the model modifies its own code directly.

Early adopters are already putting it into production. Neuroscientist Jérôme Lecoq at the Allen Institute used roughly 20 custom skills to build an automated review pipeline that synthesizes thousands of papers into 100-plus-page drafts, turning what used to take two years into a matter of days. At UCSF’s Brain Tumor Center, Stephen Francis cut the runtime of comprehensive glioma germline analyses to roughly one-tenth the previous timeline. Biotech firm Manifold Bio used the workbench to rank tissue-targeting drug candidates against internal proprietary assays.

Anthropic is offering discounted Team seats for academic labs and rolling out grants for up to 50 AI for Science projects, offering up to $30,000 in Anthropic credits and $2,000 in Modal compute, with applications open through July 15, 2026.

Why it matters

Anthropic is making a very clear strategic bet here: workflow beats raw specialty models.

Look at how the big three are dividing the life sciences market. OpenAI launched GPT-Rosalind in April as a fine-tuned biology model, but locked it behind heavy enterprise qualification and safety reviews. Google DeepMind relies on proprietary crown jewels like AlphaFold and AlphaGenome that competitors can only call as external tools. Anthropic is taking the playbook that worked for Claude Code and applying it to biology: give everyday Pro and Max users a practical operating layer that glues existing tools together.

We think this pragmatic approach gives Anthropic an immediate edge with working scientists who are tired of switching between terminal windows and web portals. Being able to run on local HPC clusters without leaking sensitive data solves a massive compliance headache.

However, there is an obvious catch. The reviewer agent tasked with validating citations and math is still the same underlying model grading its own homework. Until labs see independent validation pipelines catch edge cases, researchers will still have to double-check every reference manually.