Daily News · 1 min read

Pydantic AI Updates: May 20, 2026

pydantic-ai release gemini openai mcp

1. Pydantic AI v1.99.0 adds Gemini 3.5 Flash and fixes OpenAI strict schema regex lookarounds

Pydantic AI. v1.99.0 wires up support for Google’s new gemini-3.5-flash model so it can be selected directly through the google: provider prefix alongside the rest of the Gemini family. The release also patches an OpenAI strict-schemas bug where regex patterns containing lookarounds were being rejected, so output schemas using those patterns now serialize cleanly into Responses API requests. Source

2. v1.98.0 ships OpenAI Responses token counting and a unified Agent retries parameter

Pydantic AI. v1.98.0 adds OpenAIResponsesModel.count_tokens, giving callers a way to estimate prompt costs against the Responses API before dispatching a run. The release also collapses the Agent constructor’s separate tool_retries and output_retries arguments into a single retries: int | AgentRetries parameter, fixes the MCP integration so it no longer requires fastmcp.server at runtime, and continues the v2 staging by deprecating the pydantic_ai.ext.aci module. Source