Inspect is a framework for evaluating large language models, created by the UK AI Security Institute. It gives researchers and engineers a rigorous, reusable way to measure what models can actually do: reasoning, coding, tool use, multi-turn dialogue, and safety-relevant behavior, with scores they can defend.
Inspect is developed in the open on GitHub, where it has earned more than 3,000 stars under the MIT license, and it is one of the most actively maintained evaluation tools anywhere, with updates landing in October 2026. Evaluations are composed from three building blocks: datasets, solvers (the strategies that generate model answers, from simple generation to complex agent scaffolds), and scorers (from exact matching to model-graded judgments). The framework ships with facilities for prompt engineering, tool usage, multi-turn dialog, and model-graded evaluation, and it includes a collection of over 200 pre-built evaluations ready to run on any model, covering reasoning benchmarks, coding tasks, and safety suites. Extensions provided by other Python packages can add new elicitation and scoring techniques, so the ecosystem grows without forking the core. Because it comes from a national AI safety institute and is used for serious model assessment, Inspect emphasizes reproducibility: runs are logged in detail so results can be inspected, compared, and audited later. It works with all major model providers as well as local models.
Inspect is completely free and open source under the MIT license. Your only costs are the model API calls (or compute) that the evaluations themselves consume.
Inspect is for AI safety researchers, evaluation engineers, model developers, and policy-adjacent technical teams who need trustworthy, reproducible measurements of model capabilities and risks rather than ad-hoc prompt testing.