Claude Science
Use Claude Science when a research team on a Claude subscription needs an auditable, code-backed workbench that keeps papers, datasets, figures, and lab compute in one session — with a reviewer agent that checks citations, calculations, and whether every figure traces back to the code that made it. It runs locally on macOS or Linux and can reach an SSH or HPC login node, so large or sensitive datasets can stay on your own infrastructure while Claude receives only the context each step needs. Skip it for Coding, RAG, or Agents work — Claude Code, Codex, and Cursor-style surfaces are the comparison set there. One caveat for evaluators: Anthropic has not named a fixed backing model for Claude Science, so the Opus- and Sonnet-class picks below are LLM Reference recommendations, not a vendor default. Available as of 2026-06-30 for Pro, Max, Team, and Enterprise plans; Team and Enterprise need an admin to switch it on.
Decision note
Claude Science is a vertical workbench surface, not just another model host. Its differentiator is workflow topology — local and remote execution, curated domain skills and connectors, artifact provenance, reviewer-agent checks, and lab-compute orchestration — rather than a benchmark score. It launched 2026-06-30 and is gated by Claude subscription and admin enablement, so confirm access before committing a team to it.
Best fit
Scientific research teams on a Claude subscription that need auditable, code-backed workflows across papers, data, figures, and lab compute — with reviewer-agent checks and local, SSH, or HPC execution.
Not for
Coding, RAG, JSON / Tool use, or general Agents workflows, where Claude Code, Codex, or Cursor-style surfaces are the right comparison set. It is not a notebook replacement and not a compliance guarantee.
Overview
AI workbench for scientists combining literature analysis, scientific artifacts, curated science skills, auditable outputs, and compute orchestration across local machines, lab infrastructure, or GPUs on demand.
Compare Claude Science with other research workbenches and domain agents on model access, curated skills, and workflow coverage.
Best LLMs for this agent
Use the current Opus-class Claude when the session needs stronger multi-step reasoning across papers, datasets, code, and artifact review. Anthropic does not name a fixed model for Claude Science in the launch post, so this is an LLM Reference recommendation rather than a vendor-stated default.
Use the current Sonnet-class Claude when cost and latency matter more than maximum reasoning depth. Anthropic states Claude Science is available to Pro, Max, Team, and Enterprise users, but does not state a fixed backing model in the launch post.
Pricing and Provider Comparison
Hosted agents expose plan pricing. Bring-your-own-key agents pass the main cost through the model provider route you configure.
Agent access tiers
| Tier | Price | Includes |
|---|---|---|
| Pro | $20/mo | Claude Science access on macOS and Linux; more usage than Free, Claude Code, Claude Cowork, Claude Design, Research, projects, connectors, and additional Claude models |
| Max | $100/mo | Claude Science access with 5x or 20x more usage than Pro, higher output limits, early access to advanced Claude features, and priority access at high-traffic times |
| Team | $25/mo | Claude Science access for teams of 5 to 150; standard seats include all Claude features plus more usage than Pro. Admin enablement required. |
| Enterprise | $20/mo | Claude Science access with seat price plus usage at API rates; admins can set user and org spend limits. Admin enablement required. |
Featured Model Specs
| Model | Lab | Released | Context | Reasoning | Openness |
|---|---|---|---|---|---|
| Claude Sonnet 5 | Anthropic | 2026-06-30 | 1m | Yes | Proprietary |
| Claude Opus 4.8 | Anthropic | 2026-05-28 | 1m | Yes | Proprietary |
Model Compatibility Matrix
| Model | Lab | Default | Access tier | Notes |
|---|---|---|---|---|
| Claude Opus 4.8 | Anthropic | Yes | Pro, Max, Team, Enterprise | LLMReference recommendation for deep scientific workflows; Anthropic has not named a fixed Claude Science backing model. Team and Enterprise admins must enable the app. |
| Claude Sonnet 5 | Anthropic | No | Pro, Max, Team, Enterprise | LLMReference balanced recommendation when latency and included subscription capacity matter more than Opus-class reasoning depth; Anthropic has not named a fixed Claude Science backing model. |
Underlying Models
Capabilities
| Feature | Supported |
|---|---|
| Chat | Yes |
| Agent mode | Yes |
| Multi-file edit | Yes |
| Autocomplete | No |
| Terminal | No |
| Code review | No |
| Git integration | No |
| Deploy integration | No |
| Local models | No |
| Bring your own key | No |
| MCP support | No |
| Offline mode | No |
| IP indemnity | No |
| SSO | Yes |
| On-prem | No |
Related Anthropic Agent Surfaces
FAQ
Should I use Claude Science or Claude Code?
Use Claude Science when the output is scientific: code-backed figures and artifacts, domain databases and skills, reviewer-agent checks on citations and calculations, and jobs that run on local, SSH, or HPC compute. For building and shipping software, Claude Code remains the default. Many teams use both — Claude Science for the research, Claude Code for the engineering.
When should I use Claude Science instead of regular Claude or a notebook?
Use regular Claude to read, summarize, or lightly analyze papers; reach for Claude Science when artifacts, code provenance, datasets, and compute jobs all need to live in one auditable session. It also does not replace your notebook — it orchestrates around notebook-, terminal-, database-, and compute-style steps, coordinating and reviewing them rather than being the editing surface.
Which Claude model does Claude Science use?
Anthropic has not publicly named a fixed backing model for Claude Science. LLM Reference recommends an Opus-class Claude (Claude Opus 4.8) for hard, multi-step science work across papers, datasets, code, and artifact review, and a Sonnet-class Claude (Claude Sonnet 5) when cost and latency matter more than maximum reasoning depth.
What data leaves my lab infrastructure?
Anthropic says large or sensitive datasets can stay on your laptop, Linux box, or HPC node, and only the context each step needs is sent to Claude. Treat that as a data-flow design, not a compliance or security guarantee — validate it against your own governance before moving regulated data.
Who can enable Claude Science, and what plans qualify?
It is available to Claude Pro, Max, Team, and Enterprise plans. On Team and Enterprise, an admin has to switch the app on before members can use it.
What compute can it run on, and who pays?
Claude Science can run locally on macOS or Linux, connect to a remote machine over SSH or an HPC login node, and use your own Modal account. It drafts a plan and asks before reaching a new resource. That compute is separate from your Claude subscription — you bring and pay for it — unless a specific Anthropic program says otherwise.
What does the reviewer agent verify?
The reviewer agent checks citations, calculations, and whether figures trace back to the code that produced them. It is a provenance and consistency check, not a substitute for scientific peer review — keep human and domain validation in the loop.