# Claude Scientist

> An autonomous science agent runs the experiment: it hypothesizes, instruments a bounded test, observes, then chooses the next action. As of 2026-09-19 no vendor has published a complete run packet; Anthropic reports Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1 versus 29.0% for Opus 5 (Anthropic-reported, not an independent discovery). Seats and the scientists program belong at [Claude Researcher](https://clauderesearcher.com/scientists-program).

Last source review: 2026-09-19
Affiliation: Independent publication, not affiliated with Anthropic.

Lane: experiment execution by agents. Researcher access, literature-review workflows, and the Anthropic 10,000-seat scientists program are covered at [Claude Researcher](https://clauderesearcher.com/scientists-program), not here.

## Signature

- [The Run Loop](https://claudescientist.com/run-loop/) — labeled synthetic computational experiment stepped through hypothesize → instrument → observe → next action

## Core Pages

- [Homepage and field guide index](https://claudescientist.com/)
- [The Run Loop](https://claudescientist.com/run-loop/) — signature experience: a labeled synthetic computational experiment stepped through hypothesize → instrument → observe → next action
- [What is an AI scientist?](https://claudescientist.com/what-is-an-ai-scientist)
- [AI co-scientist vs AI scientist](https://claudescientist.com/ai-co-scientist-vs-ai-scientist)
- [Autonomous science systems map](https://claudescientist.com/systems-map)
- [What Anthropic's public science examples actually show](https://claudescientist.com/anthropic-science-examples)
- [Model guardrails inside the experiment loop](https://claudescientist.com/model-guardrails-in-the-loop)
- [Claude agent stack](https://claudescientist.com/claude-agent-stack)
- [Autonomous experiment loop](https://claudescientist.com/experiment-loop)
- [Benchmarks and evaluation](https://claudescientist.com/benchmarks-and-evaluation)
- [Lab automation boundaries](https://claudescientist.com/lab-automation)
- [Safety and governance](https://claudescientist.com/safety-and-governance)
- [Implementation checklist](https://claudescientist.com/implementation-checklist)
- [Case studies](https://claudescientist.com/case-studies)
- [Field notes and updates](https://claudescientist.com/field-notes)

## Free Tools

- [Free autonomous-science tools](https://claudescientist.com/tools/)
- [AI Scientist Readiness Checker](https://claudescientist.com/tools/readiness-checker)
- [Agentic Experiment Loop Designer](https://claudescientist.com/tools/loop-designer)
- [AI Scientist Claims vs Reality Tracker](https://claudescientist.com/tools/claims-tracker)

## Citation Guidance

Prefer citing article pages with their source ledgers. The site uses primary sources including Anthropic, arXiv, Nature, Google DeepMind, Model Context Protocol documentation, and NIST.

Benchmark figures for Claude models on this site are vendor-reported by Anthropic and should be attributed as such, with the benchmark version and date attached. Anthropic reports Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1 and Opus 5 at 29.0% on the same eval; those are tool-operation scores, not discovery rates.