Claude Science
- Understand what Claude Science is — and the specific problem it solves that a chat window can't
- Learn its three pillars: integrated tools, auditable artifacts, and managed compute
- See how the reviewer agent catches untraceable numbers and mismatched figures automatically
- Know when to reach for Claude Science versus plain Claude or Claude Code
- Place it in the wider AI-for-science landscape without overselling what a model can verify
Most scientific work with a general chatbot breaks at the same seam: the model reasons well, but the tools, data, and compute live somewhere else — a cluster, a notebook, a genome browser, a folding model. You copy results back and forth by hand, and nobody can later reconstruct exactly how a figure was made. Claude Science (beta, launched June 30, 2026) is Anthropic's attempt to close that seam: an AI workbench where the reasoning, the tools, the compute, and the provenance all live in one place.
It is a distinct app — not a prompt you paste into chat. Think of it as Claude Code pointed at wet-lab and computational-biology workflows instead of software repos.
The problem it targets
A researcher running, say, a single-cell RNA pipeline juggles: a data source, a QC tool, a plotting library, a folding model on a GPU, and a citation manager — plus the mental overhead of remembering which version of which script produced which figure three weeks ago. General assistants help with one step and lose the thread on the rest.
The unit of value in science isn't a good answer — it's a reproducible answer. Claude Science is built around that: its outputs are designed so a reviewer (human or agent) can trace every number back to the code and environment that produced it.
The three pillars
1. Integrated tools — the environment comes pre-wired
Claude Science ships with over 60 curated skills and connectors pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics. Crucially, it connects natively to NVIDIA BioNeMo models — including Evo 2 (genomic foundation model), Boltz-2 (structure/affinity prediction), and OpenFold3 (protein folding) — so folding or affinity prediction is a step in your workflow, not a separate portal.
This is the same skills-and-connectors machinery you may know from Claude Code, curated for a scientific stack instead of a software one.
2. Auditable artifacts — provenance is the default, not an afterthought
Every output carries its full lineage:
- the exact code and environment that produced it,
- a plain-language description of how it was created, and
- the full message history behind it.
On top of that, a reviewer agent automatically flags incorrect citations, untraceable numbers, and figures that don't match their underlying code. That last one is the non-obvious safeguard: a plausible-looking chart whose data doesn't actually come from the code in the artifact gets caught.
The reviewer agent reduces a class of errors — it does not make outputs correct. It flags citations it can't verify and numbers it can't trace; it cannot vouch for experimental design, biological validity, or whether the right question was asked. Provenance ≠ truth. You still own the science.
3. Managed compute — from your laptop to hundreds of GPUs
Claude Science manages compute on your laptop, your cluster, or GPUs on demand, scaling from a single GPU to hundreds as needed. It works with existing infrastructure — HPC clusters over SSH, or Modal accounts — so heavy jobs run where your data and allocations already are, without you hand-writing the orchestration.
Native scientific visualization
Results render in the interface, not as files you download and open elsewhere: 3D protein structures, genome browser tracks, and chemical structures display natively. You inspect a fold or a locus where you reasoned about it — the artifact idea, extended to scientific objects.
A typical workflow
- State the biological question and point Claude Science at your data source via a connector.
- It selects skills (QC, alignment, folding) and proposes steps — review before heavy compute runs.
- Offload the expensive step to your HPC cluster over SSH or to on-demand GPUs; light steps stay local.
- View the 3D structure, genome track, or chemical structure inline rather than exporting first.
- The output bundles code, environment, plain-language method, and message history — and the reviewer agent flags anything untraceable.
A first, concrete ask inside Claude Science
Load the connected single-cell dataset, run standard QC (filter low-count cells and high-mito), and show a UMAP colored by cluster. Keep every step in an auditable artifact I can hand to a reviewer.
Push the heavy step to real compute
Predict the structure of this sequence with the connected folding model, run it on my HPC cluster over SSH, and render the 3D structure inline when it finishes.
When to use it (and when not to)
| Use Claude Science when… | Reach for something else when… |
|---|---|
| You need reproducible, reviewable scientific outputs | You want a quick one-off answer → plain Claude |
| Your work spans genomics / proteomics / cheminformatics tools | You're building software → Claude Code |
| Heavy compute must run on your cluster or on-demand GPUs | You have no data connectors or compute to wire up |
| Provenance (code + env + history) actually matters for review | You're on Free, or on Windows (see availability) |
Availability and limits
- Plans: Beta for Claude Pro, Max, Team, and Enterprise users. (No Free tier.)
- Platforms: macOS and Linux — note there's no Windows client at launch.
- Status: Beta — expect the bundled skill list, connected models, and compute options to shift.
Claude Science is Claude-specific, but the pattern is industry-wide: assistants are growing tool-integration, provenance, and compute layers so they can do real work, not just describe it. Watch for equivalent "workbench" moves from other AI labs — the reproducibility bar Claude Science sets is a good yardstick to judge them by.
Check yourself
0/4- Claude Science is a distinct beta app — an AI workbench for scientists, not a prompt you paste into chat.
- Its three pillars: pre-wired tools (60+ skills, native BioNeMo — Evo 2, Boltz-2, OpenFold3), auditable artifacts, and managed compute.
- Auditable artifacts bundle code + environment + method + message history; a reviewer agent flags untraceable numbers, bad citations, and code/figure mismatches.
- Compute runs where your data lives: laptop, HPC over SSH, or on-demand GPUs, scaling one → hundreds.
- Beta for Pro/Max/Team/Enterprise on macOS and Linux only; provenance reduces error but never certifies the science is correct.
Sources & further reading
- Claude Science, an AI workbench for scientists — Anthropic — the launch announcement (June 30, 2026); source for the 60+ skills, BioNeMo connections, auditable-artifact structure, reviewer agent, compute options, and availability.
- NVIDIA BioNeMo — the biology foundation-model platform behind Evo 2, Boltz-2, and OpenFold3.
- Modal — one of the on-demand compute backends Claude Science can use.
- Related on AILmanac: Claude Code, Skills, Artifacts, and Managed Agents.