What changed
Pydantic AI shipped v2.6.0 on July 8, 2026 and followed it with v2.7.0 on July 9. The bigger change is in v2.6.0: the framework now records time-to-first-token for streaming model requests, adds files support to CodeExecutionTool for Anthropic and OpenAI, and refreshes Bedrock model profiles for providers including Writer, Z.AI, and Moonshot AI. The follow-up v2.7.0 release adds Azure Responses shorthand and support for xAI's Grok 4.5 model.
Why it matters
This is a product story about operational ergonomics rather than benchmark theater. Time-to-first-token is one of the clearest measurements of perceived speed in chat and agent products, so exposing it directly inside the framework makes performance tuning easier. File-aware code execution is also a real workflow upgrade: agents that generate or inspect artifacts often need to pass files through tools instead of squeezing everything through plain text.
Builder takeaway
The practical pattern is that agent frameworks are competing on production friction now. Pydantic AI's latest releases make it easier to watch latency, wire richer tool inputs, and test newer provider options without custom glue. For builders shipping portfolio projects, internal tools, or customer-facing agents, that kind of integration work often matters more than one more headline model launch.