Anthropic just released Claude Science, a workbench aimed at people doing real computational research. Read the original announcement here: anthropic.com/news/claude-science-ai-workbench.
First thing to get straight, because most of the coverage buries it: this is not a new AI model. It runs the same Claude models you already have. What's new is the environment around them. If you've used Claude Code, you already know the shape of this. Claude Code took the model and wired it into your terminal, your files, and your git history. Claude Science does the same thing for a scientist's stack: the databases, the notebooks, the compute, the file formats a researcher actually works in.
The problem it's aiming at
Computational research is a tab-management job. You pull sequences from one database, run a pipeline in a second tool, push the job to a cluster you SSH into from a terminal, then drag the output into a notebook to make the figure. Every handoff is a place to lose track of what you did. Six months later a reviewer asks how you generated a plot and you're digging through shell history trying to reconstruct it.
The pitch is one environment instead of ten. The databases, the pipelines, and the compute all sit behind the same chat window, and the model drives them.
What's actually in it
- 60+ scientific databases wired in. UniProt, PDB, Ensembl, ChEMBL, ClinVar, GEO and more, so Claude can pull real data without you writing the fetch code. It ships pre-configured for genomics, single-cell, proteomics, and cheminformatics.
- Native rendering of scientific artifacts.3D protein structures, chemical structures, and genome browser tracks show up in the workbench itself. You're not exporting a file to open it in a separate viewer.
- Reproducibility built into every result.When Claude generates a figure, it attaches the exact code and environment that produced it, a plain-language description of how it was made, and the full message history. That's the part that matters. Every result traces back to the code that made it.
- Compute it can actually manage. This is the standout. Claude Science can SSH into an HPC cluster, scale a job from a single GPU to hundreds as the work demands, and manage those resources for longer training runs. It drafts a plan, asks before it reaches for new resources, and lets you review or revoke any decision before it submits the job.
Why the compute piece is the interesting one
A chat window that reads a protein database is a nice demo. A chat window that provisions and manages HPC resources is a different category of tool. The gap between "I know which analysis to run" and "the job is queued on the cluster with the right resources" is exactly the part that eats a researcher's afternoon, and it's the part that has nothing to do with the science.
The guardrail design is the right call here too. It plans, asks before touching new resources, and lets you revoke a decision before the job runs. You do not want a model silently spinning up hundreds of GPUs on your grant budget, and Anthropic clearly knew that.
Who this is for
This is built for people doing genuine computational research: bioinformaticians, computational chemists, genomics and proteomics researchers, anyone whose day is already a chain of databases, pipelines, and cluster jobs. If that's your workflow, this collapses a lot of context-switching into one place, and the reproducibility trail solves a real problem you already have.
Who it's not for
If you're not doing computational science, this isn't a general research assistant with a nicer coat of paint. The value is entirely in the domain integrations: the databases, the artifact rendering, the HPC access. Strip those out and you're back to a normal Claude chat, which you can already get. It also assumes you know what analysis you want to run. It removes the plumbing, not the expertise. A tool that SSHes into your cluster is only as safe as your judgment about what to let it run, so the review-and-revoke step isn't optional, it's the whole safety model.
How to get it
Claude Science is in beta for Pro, Max, Team, and Enterprise users. On Team and Enterprise plans an admin has to switch it on for the organization first. Same underlying models as always, so there's no separate access to request. If you're on one of those plans and your work is computational, it's worth an afternoon to see whether it fits your stack.