API discovery: Pydantic AI has added an openai-codex provider for a ChatGPT/Codex subscription. After the official Codex login, the public guide’s smallest model selection is simply Agent('openai-codex:gpt-5.6-luna'). That is useful because it puts subscription authentication behind a named provider boundary instead of asking every typed-agent application to improvise its own OAuth bridge—or to reach for a pay-per-token key it does not need.
Try it
Use it for a deliberately small local proof of concept: give a one-turn Pydantic AI agent a Markdown release note and ask it to return a short “what changes for an operator?” brief. Keep the task read-only and the output disposable. The point is to confirm that your project’s model/provider wiring is correct before you add tools, background work, or a persistent credential source.
Copy-paste agent instruction
Set up a small Pydantic AI proof of concept using the openai-codex provider. First confirm that I have completed the official Codex login; do not read, print, commit, or upload any credential files. Create a one-turn, read-only agent that turns a local Markdown release note into a short operator-facing changelog. Report provider errors clearly and do not retry authentication blindly.Test caveat
The guide is candid about the boundary. This requires an eligible subscription plus CLI or browser OAuth login; the provider is streaming-only, forces store=false, has no device flow, and does not provide token counts through this authentication path. Its public test uses a recorded authenticated fixture and checks that credential values stay out of tracing, but this reporting run did not sign in or execute a live request. Treat account access, refresh credentials, model availability, and cost behavior as things to verify in your own controlled environment.