Models and providers

attest grades with Anthropic, OpenAI or Gemini models. Pick one when you create the Attest grader:

from attest import Attest

Attest()                                        # Anthropic, default model
Attest(provider="openai", model="gpt-6.1-terra")
Attest(provider="gemini")
Attest.models("openai")                         # models you can use

On the command line, pass --provider and --model to any command that grades with a model.

Provider	Install	Key
anthropic	included	ANTHROPIC_API_KEY
openai	pip install "agent-attest[openai]"	OPENAI_API_KEY
gemini	pip install "agent-attest[gemini]"	GEMINI_API_KEY or GOOGLE_API_KEY

Keys are read from the environment or from a .env file in the working directory or a parent. You can also pass the chosen provider's key as key= to Attest.

Each provider defaults to a small, fast model, so grading a run costs cents. A stronger model catches subtler errors. Compare models on your own labeled runs before you switch, and use difference_is_real to check that the gap isn't noise.

Your own client

Pass any object that implements the LLMClient protocol in attest.llm as client=, to route grading through your own gateway, cache or rate limiter.