What happened
Pydantic released Pydantic AI v2.62.0 on GitHub, with the release page dated October 26, 2026 and a GitHub release timestamp of October 27. The release’s most visible model-control change is the ability to change thinking effort in the middle of a conversation. It does this through OpenAI’s configuration_update and Anthropic’s per-message output_config, while recording the change as a ModelSettingsDeltaPart. The notes also say that Pydantic AI preserves the prompt cache when changing thinking effort mid-conversation on OpenRouter, using OpenRouter’s configuration_update system message.
The release introduces three model error categories: ModelRequestRejectedError, ModelAuthenticationError, and ModelNotFoundError. It also changes hosted MCP capability tool naming: tools are prefixed with the service by default, and developers can configure that behavior through tool_prefix. Related MCP changes include appending changed server instructions with on_change='append' and assigning an id-less URL MCPToolset its hostname as an id.
Several changes concern how an agent preserves and presents state. A SummarizingCompaction summary is now stored as a CompactionPart rather than a system prompt, and provider-less CompactionPart objects are sent to every model. The framework can let a model request compaction at a natural break through CompactOnRequest; it also adds AnthropicModel.compact_messages() for on-demand native compaction. Loaded deferred-capability instructions are stated in the system voice, their changes can be appended, and they are exposed to before_model_request. Harness Memory snapshots are recorded in history once per changed run so requests remain append-only.
The release adds other execution and integration capabilities. It supports provider batch APIs with a BatchModelRequests capability and a get_wrapper_model capability hook. In Temporal durability, tool calls can run as child workflows using child_workflow tool metadata. SubAgents can run delegations as tasks that a model can background, steer, answer, and cancel when task_control is enabled; Coder(task_control=) enables that control. CodeMode(monty_sandbox=...) accepts a remote Monty URL or a function that builds the pool, while monty_sandbox_url is deprecated.
There are compatibility notes worth treating as operational changes. GraphBuilder broadcast branches now start in .to() order rather than reversed order; the release warns that this can break replay of durable workflows paused across an upgrade. The notes also specify behavior for realtime shared settings: they can warn, raise, or ignore like standard ModelSettings; forced tool_choice raises on Gemini Live, while parallel_tool_calls=False warns. Other fixes include reporting failed tool calls to realtime models so they answer rather than go silent, and closing event streams and in-flight tool calls correctly when consumers or listeners exit early.
What the archive adds
This is a new release in a known Pydantic AI thread, not a restatement. The immediately preceding Pydantic AI v2.61.0 Field Note covered structured citations, Kimi K3.1 Pro model names, and guardrail fixes. Version 2.62.0 shifts the emphasis toward runtime model controls, error classification, MCP namespace behavior, and workflow durability.
Trend signal
The release supports Agent protocols become the interoperability layer. Prefixing hosted MCP tools with their service is a small but practical response to a real integration boundary: once multiple remote services contribute tools to one agent, their names must remain distinguishable and configurable. The accompanying handling of changing MCP instructions further treats the protocol connection as live application state, rather than a static prompt addition.
Why it matters
Teams often experience agent reliability as a collection of loosely connected problems: a model is too expensive on one turn, unavailable on another, a remote tool collides with a similarly named one, or a resumed workflow no longer behaves as it did before an upgrade. This release exposes controls and state representations for those problems inside the framework.
The mid-conversation thinking-effort control is not a claim that the framework can automatically choose the right reasoning budget. It does give an application a documented mechanism to make that choice while a conversation is running. The new exception categories similarly do not solve provider failures, but they enable a harness to distinguish authentication remediation, model selection remediation, and request rejection. For MCP-based systems, service-prefixed tool names create a more legible default when capabilities from multiple hosted services coexist.
Limits and uncertainty
The release notes describe APIs and behavior changes, not measured production outcomes. They do not establish how much changing thinking effort affects quality, latency, or cost for any task, nor do they provide migration time or failure-rate data for the new error types. The notes say service prefixes are the default for hosted MCP capabilities, but they do not establish a universal naming convention across MCP implementations. The GraphBuilder warning identifies replay risk for paused durable workflows, but does not enumerate every affected workflow shape. Teams should test their own provider combinations, tool registries, cached conversations, and replay paths before treating this as a drop-in upgrade.
Skill to build now
Add a release-upgrade test that restores one paused durable workflow and runs it through its next tool call. In the same test environment, inspect the final tool names after registering each hosted MCP capability, then confirm that application routing and allowlists use the resulting names. Separately, map the three new model errors to explicit application actions: credential remediation for authentication errors, fallback or configuration checks for a missing model, and a user- or policy-facing path for rejected requests. This turns the release’s new types into observable operational behavior.
Reader takeaway
Pydantic AI v2.62.0 is less about a new model and more about making agent runtime state explicit: reasoning settings can change by turn, provider failures have distinct types, and hosted MCP tools gain clearer namespaces. The upgrade is most valuable to teams willing to test it against their real conversation, tool, and workflow-replay paths.
Open question
Will Pydantic AI users adopt service-prefixed MCP names as a stable application-level contract, or treat them as a configurable display detail that must be normalized elsewhere?