Anthropic Puts Claude on the Lab Bench With Claude Science

Anthropic wants to move Claude out of the chat window and onto the lab bench. On 1 July the company launched Claude Science, a workbench aimed at researchers that pulls databases, coding notebooks, terminals and specialist tools into one environment. It is not a new model. It is an attempt to fit the model to the messy, multi-step reality of how scientific work actually gets done.

The pitch is integration. Claude Science connects to more than 60 databases and specialist agents, runs locally on macOS and Linux or remotely over SSH and HPC clusters, and routes heavy computational jobs to existing high-performance computing accounts or to Modal, according to TechFundingNews. It also leans on NVIDIA's BioNeMo Agent Toolkit and specialised biology models such as Evo 2, Boltz-2 and OpenFold3, which points squarely at drug discovery as the first serious use case.

Built to be checked

The detail worth dwelling on is auditability. Anthropic says the workbench produces outputs that carry a code trail alongside a plain-language explanation of what was done. For a working scientist that matters more than raw capability. A result you cannot reproduce or defend is not much use in a paper, and the persistent complaint about using general assistants for research has been that they hand back confident answers with no way to trace the reasoning.

Anthropic is not pitching this from the sidelines. The company says it has been running Claude Science internally for pre-clinical drug discovery aimed at neglected tropical diseases. Michael Pollastri, a chemist involved in that work, said the tool could "genuinely speed" the job of identifying candidate compounds by automating the slow business of gathering and cross-referencing information.

A three-way race

Anthropic is late to a space its rivals reached first. Google introduced a "Gemini for Science" workbench at its May I/O conference, bundling in more than 30 life-science databases, and OpenAI released a research-focused system called GPT-Rosalind in April aimed at drug discovery. What was a quiet corner of the market a year ago is now a contested one, with all three labs trying to sit next to scientists rather than sell to them from a distance.

To seed adoption, Anthropic is backing up to 50 "AI for Science" projects with as much as $30,000 in credits each, with applications open through 15 July. Access to the workbench itself is in beta for Claude Pro, Max, Team and Enterprise subscribers.

The interesting question is whether any of this changes the pace of real research or simply moves familiar tools behind a single login. We have already covered cases where AI genuinely earned its place in the lab, from an AI chemist improving a stubborn drug-making reaction to models helping doctors crack rare disease cases. A workbench that makes those wins routine, and reproducible, would be worth the attention. One that mostly rebrands existing pipelines would not.