Pydantic AI v2.58.0

Summary

Pydantic AI v2.58.0 adds run-level and request-level timeouts, steering for delegated tasks, a provider that lets an app's users sign in with their ChatGPT plan, and an ElevenLabs realtime model. The release title says 2026-10-19, and GitHub shows it published on 2026-10-20.

Key points

• Timeouts: agent.run(..., timeout=...) and RunContext.deadline bound a whole run. ModelSettings['request_timeout'] limits one model request, including retries and stream consumption. ModelSettings['max_retries'] sets the provider SDK's retry count. A new UnsupportedSettingWarning fires when a model ignores a setting.
• Steering: DelegationTasks.steer redirects a delegated task while it runs.
• openai-chatgpt provider: users sign in with ChatGPT and use their own plan instead of an API key.
• Realtime: ElevenLabsAgentsRealtimeModel runs on ElevenLabs Agents, and reasoning_summary is available for GPT-Live delegation.
• Durable runs: the SQLite, Postgres, and Mongo step stores save each message once instead of repeating history in every snapshot, which cuts storage.
• Skills harness: Skills(..., unknown_names='skip') ignores include or exclude names that no library provides.
• Upgrade with pip install -U pydantic-ai.

Why it matters

Hard deadlines on agent runs and model requests are a common production need that used to require wrapper code. The ChatGPT sign-in provider lets apps run on users' own plans instead of the developer's API bill.