The release adds structured citations to TextPart and normalizes citation data from Amazon Bedrock, OpenRouter, and xAI. Citations can be replayed in message history to the provider that produced them, and the Vercel AI adapter retains them.
The version also adds the kimi-k3.1-pro and kimi-k3.1-pro-highspeed model names, async authentication functions for hosted MCP capabilities, and Shell.max_log_bytes for bounding background and shell-job logs. Pydantic AI made parts of its durable-engine extension surface public, including operation types, ParameterTransports, DurableRunContext, and run identity hooks.
Compatibility and fix notes include warnings for provider-specific thinking settings overridden from a broader scope. For models that cannot disable thinking, Pydantic AI sends the least thinking level accepted when thinking=False. InputGuardrail now redacts the user's own prompt and closes fail-open paths. The release notes do not establish citation accuracy, provider-wide behavior beyond the listed integrations, migration effort, or the effect of these guardrail changes in a particular deployment.