Pydantic AI Updates: October 27, 2026

1. Pydantic AI 2.62.0 Adds Provider Batch APIs and Mid-Conversation Thinking Effort Changes

Pydantic AI released v2.62.0. A new BatchModelRequests capability adds support for provider batch APIs. Thinking effort can now change mid-conversation through OpenAI’s configuration_update and Anthropic’s per-message output_config, recorded as a ModelSettingsDeltaPart, and on OpenRouter the prompt cache is kept. The release adds ModelRequestRejectedError, ModelAuthenticationError and ModelNotFoundError error categories, plus AnthropicModel.compact_messages() for on-demand native compaction. allowed_tools on OpenAI Chat Completions restricts tool choice while preserving the prompt cache. TemporalDurability can run tool calls as Temporal child workflows. Several changes carry compatibility notes. Hosted MCP tool names are now prefixed with the service by default, SummarizingCompaction summaries are stored as CompactionParts, and GraphBuilder broadcast branches now start in .to() order, which can break replay of durable workflows paused across the upgrade. Source

2. SubAgents Gain Task Control, and Harnesses Can Compact on Request

Pydantic AI v2.62.0 also extends its harness tooling. SubAgents can run delegations as tasks that the model can background, steer, answer and cancel, behind a task_control option that Coder(task_control=) also exposes. A new CompactOnRequest capability lets the model ask for a harness compaction at a natural break in a task. CodeMode(monty_sandbox=...) now accepts a remote Monty URL or a pool-building function, replacing the deprecated monty_sandbox_url. The release also adds StructuredContent user content and uses the Pydantic AI Gateway’s reported cost for RequestUsage.cost. Source