Anthropic just gave scientists a single place to run their entire research workflow, and the clever part is that it isn’t a new model at all.
On June 30, 2026, Anthropic launched Claude Science, an AI workbench built for researchers. The pitch is simple: instead of bouncing between PubMed, Jupyter, R, a cluster terminal, and a dozen databases that each speak their own language, you do all of it in one environment while an AI agent handles the messy plumbing in between.
Here’s the part most coverage buried. Claude Science is not a new AI model, and it is not a smarter model for biology. It runs the exact same Claude everyone already pays for, Opus 4.8 included, with no special access and no gating. The bet is on the workflow around the model, not the model itself. Let’s break down what that actually means, what scientists are doing with it, and how it stacks up against what OpenAI and Google are shipping.
What Claude Science actually is
Think of it as Claude Cowork for the lab. It is a dedicated app where a coordinating agent moves across your literature, your code, your data, and your compute, and produces figures and manuscripts you can trace all the way back to the code that made them.
According to Anthropic’s announcement, you interact with a generalist agent that has access to over 60 curated skills and connectors, pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics. That agent can spin up specialist sub-agents, hand tasks to custom agents you build yourself, and run a separate reviewer agent that checks citations and calculations, flagging and correcting errors before they reach your paper.
It runs wherever you already work: locally on macOS or Linux, or on a remote machine over SSH or an HPC login node. This is the same design philosophy behind Claude Skills, where you teach the tool your process once and it applies it every time after.
The whole thing builds on Claude for Life Sciences, which Anthropic launched back in October 2025 to make the regular Claude chatbot better at research tasks. Claude Science is the dedicated home for that work.
The three things that make it useful
1. Every figure is reproducible by default
Science is visual, so Claude Science renders rich artifacts natively: 3D protein structures, genome browser tracks, chemical structures, and more. But the useful bit is what ships alongside each figure.
When it generates a figure, it includes the exact code and environment that produced it, a plain-language description of how it was made, and the full message history. You can come back six months later and understand precisely how a result was reached, which makes validating and reproducing it far easier. Want to remove gridlines or switch an axis to log scale? Ask in plain English and the agent edits its own code.
2. It manages your compute and scales on demand
Big analyses like folding a protein or running a genomics pipeline usually mean setting up a job, shipping it to a cluster, waiting, checking if it failed, and pulling the results back. Claude Science handles that loop for you.
It drafts a plan, asks before reaching any new resource, and lets you review or revoke a decision before it submits a job to the infrastructure your lab already uses, whether that’s your own HPC cluster over SSH or a Modal account for compute on demand. It scales from a single GPU to hundreds as needed. Because it runs on your own machines, large or sensitive datasets never have to leave the systems they already live on, and only the context needed for each step gets sent to Claude. You can also fork a session at any point to test two approaches without losing the original thread.
3. It is domain-ready on day one
Biology data is scattered across dozens of specialized sources, each with its own schema and query language: UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO, plus journals and preprint servers. Ask Claude Science a plain-language question and specialist agents query and synthesize across all of them so you don’t have to navigate each one by hand.
It also connects natively to NVIDIA’s BioNeMo models and libraries, including Evo 2, Boltz-2, and OpenFold3. Already have pipelines and tools you trust? Save any pipeline as a reusable skill or connect your lab’s preferred tool through a connector, and future sessions inherit it automatically.
What scientists are already doing with it
Over the past few months, beta users have run single-cell RNA sequencing analysis, CRISPR screen design, protein structure prediction, and cheminformatics. Three examples stand out.
Manifold Bio designs tissue-targeting medicines that home to a specific organ or cell type so a drug acts where it’s needed and spares the rest of the body. The company used Claude Science to nominate targets for its latest experiments, assessing surface expression, trafficking, and safety for each tissue and target, then ranking candidates against criteria learned from its own proprietary data. What set it apart from a general coding assistant, Manifold said, was that it could run the task end to end with the context of past programs built in.
Jérôme Lecoq, a neuroscientist at the Allen Institute, built a multi-agent “computational review template” of about 20 custom skills for writing long-form reviews. Sub-agents read through thousands of papers, pulled the central claim and key finding from each, and stored them in an evidence database. The pipeline then wrote the review section by section, using actor-critic pairs where one agent creates and a separate reviewer checks for accuracy and citation fidelity. A review that used to take his team as long as two years? He now has about 10 of them, many over 100 pages, with citations checked by reviewer agents.
Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, used it to support studies on the molecular epidemiology of glioma. His lab reports that Claude Science ran comprehensive germline workups across multiple approaches in roughly one-tenth the time it used to take. His group independently validated the results, confirming the analyses were both fast and robust.
How it’s different from OpenAI and Google
This is where the “not a new model” decision gets interesting, because Anthropic’s two biggest rivals went the opposite way.
OpenAI’s GPT-Rosalind, launched in April 2026, is a purpose-built model fine-tuned specifically for biology, drug discovery, and genomics. It’s powerful, but it’s locked behind a trusted-access program for vetted Enterprise customers, and even after a June expansion it stays limited to qualified research organizations. Google’s DeepMind, meanwhile, keeps advancing specialized standalone models like AlphaFold and AlphaGenome that each solve one slice of the problem brilliantly.
Claude Science takes a different route. Here’s the contrast in plain terms:
| Claude Science | GPT-Rosalind | |
|---|---|---|
| What it is | A workbench (harness) on existing models | A new, specialized biology model |
| Who can use it | Any Pro, Max, Team, or Enterprise subscriber | Vetted Enterprise research orgs only |
| Where data runs | Your own laptop, cluster, or HPC node | Enterprise cloud deployment |
| Core bet | Workflow, reproducibility, and integration | Deeper domain reasoning in the model |
The strategic read is that Anthropic wants to own the operating layer for science the way Claude Code became the operating layer for software. It isn’t racing to build the smartest science brain. It’s wrapping the brain everyone already has in the scaffolding a working lab actually needs.
How to get started
- Check your plan. Claude Science is in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans. On Team and Enterprise, your admin has to switch it on first.
- Install and open the app from claude.com/science. You do not have to be a scientist to try it.
- Point it at where you work. Run it locally, or connect it to your cluster over SSH or your Modal account for on-demand compute.
- Ask a real question in plain language. Let the specialist agents pull from the 60-plus connected databases instead of querying each one yourself.
- Save what works as a skill. Turn any pipeline you rely on into a reusable skill so future sessions inherit it automatically.
- Always check the reviewer’s work. The reviewer agent flags bad citations and mismatched numbers, but it runs on the same model, so treat it as a first pass, not a final verdict.
Running an academic or nonprofit lab? Anthropic now offers a Team plan with discounted seats for active research labs. And through its AI for Science program, it’s funding up to 50 projects with up to $30,000 in credits each, with Modal adding up to $2,000 in compute for selected projects. Applications are open through July 15, 2026, awards go out by July 31, and projects run from September 1 to December 1, 2026.
Honest caveats worth knowing
- It’s beta, and macOS or Linux only. No Windows support yet, and things will change as feedback comes in.
- The reviewer isn’t independent. It uses the same underlying Claude, so it can share the same blind spots. Human validation still matters.
- Biology first. The early focus is biology and biomedical research. Other fields can use it, but the pre-built depth lives in the life sciences for now.
- Faster research isn’t finished research. A January 2026 PitchBook analysis found that despite more than $17 billion invested in AI drug discovery since 2019, no AI-developed drug has yet completed a large-scale clinical trial. Speeding up the screen doesn’t guarantee a cure at the end of it.
What this really means
Once an AI can generate a figure, a pipeline, or a hundred-page review cheaply, the scarce thing stops being generation. It becomes judgment: auditing, validating, and correcting what the machine produced. That was always the actual work of good science, and Claude Science leans into it by shipping a reviewer agent and tracing every output back to its code.
The bigger question nobody can answer yet is whether tools like this speed up real discovery or just help us produce more of it faster. The labs testing it now will give the first honest answer. Either way, the strategy is worth noticing: while others build specialist brains and lock them in enterprise vaults, Anthropic bet that the workbench matters more than the model.
The smartest model in the world is useless if a scientist still has to babysit ten tools to use it. Claude Science is a bet that the plumbing was the problem all along.
Just Being Resourceful! 
Leave a comment