If you use Pydantic AI, check the compatibility notes before upgrading: thinking=False now sends the minimum thinking a model accepts, and some tool-return and graph branching behavior changed.
Key takeaways · AI-distilled
Pydantic AI v2.61.0 adds structured citations to TextPart, normalizes citations from Bedrock, OpenRouter and xAI, and replays them in message history to the provider that produced them, so source references survive multi-turn runs.
The release closes fail-open paths in guardrailsThe checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.Full definition →: InputGuardrail now redacts the user's own prompt, and also rewrites the history copy of that prompt when the model was sent one.
A step-store fix keeps dict key order when histories reload, so resumed runs keep the prompt cache.
thinking=False now sends the least thinking a model accepts on models that cannot turn thinking off, and the library warns when a provider-specific thinking setting from a broader scope overrides the thinking option.
Terms in this piece · Glossary
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
