Pydantic AI v2.61.0: structured citations

Summary

Pydantic AI 2.61.0 (2026-10-23, from the GitHub release feed) adds structured citations on text parts. Citations from Bedrock, OpenRouter, and xAI are converted to one shape, and they are sent back to the provider that produced them on later turns. The release also adds the Kimi K3.1 Pro models, a log-size limit for background shell jobs, and a public extension surface for durable execution.

Key points

• Citations: a TextPart can carry structured citations in one normalized form, whichever of Bedrock, OpenRouter, or xAI produced them. This makes it easier to show sources in a RAG or search app without code for each provider.
• Models: kimi-k3.1-pro and kimi-k3.1-pro-highspeed were added (see the Kimi K3.1 Pro note).
• Shell jobs: Shell.max_log_bytes limits the logs kept from background and shell jobs.
• Realtime: RealtimeClientSecret.webrtc_url (with webrtcfilter=on on Azure OpenAI) and RealtimeModelProfile.supports_audio_output.
• Durable execution: the durable engine's extension surface, including operation types and DurableRunContext, is now public, so third parties can write engines.
• Upgrade: pip install -U pydantic-ai.

Why it matters

Citations in one shape across providers remove a common chore in source-grounded apps. Pydantic AI supported the new Kimi model on its launch day.