Petri

What is Petri?

Open tool for automated behavioural audits of models, started by Anthropic.

Who is Petri best for?

Petri is best for AI safety researchers who probe language models for behaviors such as deception or sycophancy.

• AI safety researchers

  It turns a plain-language hypothesis into many parallel multi-turn audits and scores each transcript, so a new behavior can be tested with only minutes of hands-on work.

• Pre-release model checks

  Anthropic used automated auditing agents like it for Claude system cards, and the UK AI Security Institute built evaluations with an early version for its own model testing.

• Comparing models

  It works with the major model APIs, and Anthropic's pilot ran 111 seed instructions across 14 frontier models to compare them on the same scoring dimensions.

• Custom evaluation suites

  It ships over 170 built-in seeds and 38 judging dimensions, and lets you add your own seeds, rubrics and auditor skills on top of the Inspect AI framework.

How much does Petri cost?

Petri is free and open source under the MIT licence, with no paid tier; the only cost is the model API usage each audit runs up.

Petri runs on your own API keys for the three models it drives: an auditor that runs the conversation, the target being tested and a judge that scores it. The docs warn that a full default run over all built-in seeds takes a few hours and a non-trivial API spend, so start with one tag or a handful of seeds. Full pricing on github.com. We checked these prices on 28 Sep 2026.

How do readers rate Petri?

No one has rated Petri yet. Be the first.

Is Petri open source?

Yes. Petri's code is public under the MIT licence, so you can read it, change it and run it yourself.

Questions about Petri