Claude Science: A Beginner’s Hands-On Guide (2026)

Anthropic dropped Claude Science on June 30, 2026, and researcher timelines lit up within hours – Whitehead Institute and UCSF folks posted testimonials, and half of biotech Twitter started arguing about whether this beats OpenAI’s GPT-Rosalind. If you’re a beginner staring at all this hype, you probably have one question: do I actually need this, or is Claude Code enough?

Short answer: if you’re writing scripts, stick with Claude Code. If you’re wrangling databases, figures, and reproducibility receipts for a paper, Claude Science is the better fit – and it’s already included in the Pro subscription you might have. This guide walks you through installing it, running one real (boring) session, and the beta gotchas the launch posts skip.

Claude Code vs. Claude Science: pick one

These two products get lumped together because they look similar and share a codebase mindset. They’re not the same thing, and choosing wrong wastes a weekend.

The cleaner heuristic: Claude Science is for people whose output is a figure, a manuscript, or a defensible number. Claude Code is for people whose output is a working program. Different outputs, different tools – the surface-level similarity (both run code, both use Claude) has been confusing people since launch.

What Claude Science actually is (in plain English)

Not a new model. Anthropic’s launch post is explicit: Claude Science is “not a new AI model and not a more capable model for biology” – it runs the same Claude models already available to everyone, including Claude Opus 4.8. Think of it as a desktop app that wraps Claude in a scientist-shaped workflow.

The three pieces that matter:

• A local sandbox that writes and runs Python, R, and shell code against folders you approve – your data doesn’t leave your machine.
• Connectors to 60+ scientific tools – a generalist coordinating agent pre-configured for genomics, single-cell, proteomics, structural biology, cheminformatics, and more (per Anthropic’s announcement).
• Provenance receipts on every output – every artifact includes the exact code that generated it, the environment it ran in, a plain-language description of what was done, and the full conversation history. Results stay reproducible months later, by anyone on your team. A background reviewer also flags any claim it can’t trace to evidence before results surface (more on the reviewer’s limits in a moment).

That third piece is the actual selling point. Every AI chat tool can talk about your data. Very few hand you back a figure plus the exact code, conda environment, and message history that produced it – bundled so a reviewer six months from now can rerun the whole thing.

Install it in 10 minutes

Requirements first. Claude Science runs on macOS 13 or later and Linux x64 – that’s it. Windows users, sorry: you’re on WSL2 with a Linux install, and there’s no announced timeline for native Windows support.

You also need a paid plan. The beta covers Pro, Max, Team, and Enterprise (free tier is out). Team and Enterprise users: your admin has to enable Claude Science first, otherwise the app won’t authenticate.

macOS: Download the installer from claude.com/product/claude-science and double-click. First launch takes a few minutes while it builds Python and R environments.

Linux:

curl -fsSL https://claude.ai/install-claude-science.sh | bash

After install, a browser tab opens for sign-in. SSH’d into a remote lab machine? The OAuth redirect may never land – look for the “Paste a code” option on the sign-in screen. The docs flag it as a fallback but it’s easy to miss the first time.

Before you tidy up: Everything Claude Science installs lives in one folder – ~/.claude-science. Per Anthropic’s docs, deleting that folder wipes every project, artifact, and conversation you’ve ever run. There’s no cloud backup because Claude Science is local-first. Back up important artifacts to git or a shared drive before you clean house.

Your first session: don’t start with the flashy demo

Every launch article shows the phenylketonuria drug-candidate discovery demo. Don’t start there. Pick a task where you already know the answer – that way you catch mistakes before they matter.

Here’s a beginner walkthrough that surfaces the actual UX quirks:

1. Create a new project. The setup wizard asks which connectors to enable – for a first run, turn on bioRxiv/medRxiv and maybe ChEMBL. Leave the rest off. Fewer tools = fewer approval prompts.
2. Grant read-only access to one folder. Put a small CSV in it (50 rows of gene names or paper titles works fine). Claude Science asks before it touches new folders – approve this one as read-only.
3. Ask a boring question. Something like: “Look at genes.csv, group them by chromosome, and make a bar chart.” This exercises code execution, artifact generation, and figure rendering without any real risk.
4. Inspect the artifact panel. Click the figure. You’ll see the Python that made it, the conda env hash, and the message history. Screenshot it – that’s your proof the reproducibility loop works before you trust it with real data.
5. Ask for an edit in plain language. “Change the axis to log scale.” The agent rewrites its own code and regenerates the figure. This is where the workbench feels different from a plain ChatGPT session.

Once that clean workflow succeeds, then point it at your real data. Community consensus forming fast: never pilot a new agent on your riskiest dataset.

Three beta gotchas the launch posts skip

Beta software behaves like beta software. These three issues bit early users in the first 72 hours.

The reviewer isn’t a second scientist. This is the biggest misconception. Turns out the reviewer only checks claims against the execution record – it never re-runs the analyses themselves (confirmed in Anthropic’s overview docs). So if Claude writes buggy code that produces plausible-looking output and then correctly describes that output, the reviewer says “looks consistent” and moves on. It catches hallucinated citations and untraceable numbers, not flawed methodology. Every analysis still needs a human check before it hits a manuscript.

You’re sharing your regular Claude rate limit. Claude Science draws from the same Pro/Max plan quota – no separate allocation. As a rough guide, most people can send about 15-40 messages to Claude every five hours on Pro (per Engadget’s coverage of Claude’s rate limits, as of mid-2026). A long genomics session with lots of tool calls burns through that faster than regular chat does. If you plan to run multi-hour analyses, Max plan or Team is a more realistic tier.

Linux install can fail silently. The catch: Linux won’t start unless bubblewrap 0.8.0+ is installed and unprivileged user namespaces are enabled (check with bwrap --version). Older CentOS distros ship with namespaces disabled by default. If the app quits without a browser tab opening, that’s almost always why.

How it stacks up

Three AI-for-science strategies are colliding right now.

Claude Science goes wide: any Pro subscriber gets it (Pro runs around $20/mo as of mid-2026 – check current pricing at claude.ai). The provenance receipts are the real edge – no other tool in this space currently bundles code + environment + conversation with every figure by default.

OpenAI GPT-Rosalind (released April 2026) goes narrow: gated to enterprise customers. If you’re at a pharma company with an enterprise contract, worth evaluating. If you’re a grad student, you won’t get access yet.

Google DeepMind leans on owned models nobody else has – AlphaFold, AlphaGenome, plus Gemini for Science bundling 30+ life science databases. If your work centers on protein structure prediction specifically, DeepMind’s stack is still the reference. Claude Science can call AlphaFold via connectors, but it doesn’t own the underlying model.

Honest read: for most academic and biotech workflows, Claude Science is currently the most accessible option that produces defensible outputs. That shifts the moment OpenAI opens Rosalind access, or Google ships a real workbench layer around Gemini for Science.

What to try next

Install it today, run the boring CSV walkthrough above, then pick one real workflow you’ve been dreading – a literature review, a QC pass on RNA-seq counts, a figure you keep meaning to redraw – and hand that to Claude Science with a dataset whose answer you already know. Compare its output to what you’d have done manually. That’s the only benchmark that matters for your specific work.

FAQ

Is Claude Science free?

No. It’s included in Pro, Max, Team, and Enterprise plans – no separate charge, but the free Claude.ai tier doesn’t get it.

Can I use Claude Science for clinical or diagnostic work?

No – Anthropic is explicit that it’s not intended for clinical or diagnostic use. The reviewer agent reduces but doesn’t eliminate errors, and the whole system is designed for research workflows. If you’re at a regulated pharma company, treat every output as a draft that needs your validated pipeline to check. One researcher quoted in coverage of the launch put it well: it’s a co-pilot that requires a skilled pilot, not a shortcut to discovery.

Does my data leave my computer?

Mostly, no – files and compute stay local (or on your lab’s Slurm cluster). But prompts and Claude’s responses still travel to Anthropic under standard data retention. Anything you paste inline into the chat gets processed by their servers. If your dataset is under an IRB or data use agreement that prohibits third-party processing, talk to your admin before dropping data snippets into a Claude Science conversation.