Pydantic AI v2.56.0 adds SmolSandbox and Gemini 4 Argon support

Pydantic AI v2.56.0 was tagged on 13 October. Its new features include SmolSandbox, a workspace that runs harness capabilities in a local smolvm microVM, and a Render Workflows integration for durable execution engines. It also adds PGTaskDurability, based on pgtask.

Model support gets several updates. GoogleModel now accepts gemini-4-argon with long decode past the per-request output cap, and the default timeout for Google providers rises to 20 minutes. The release adds Python 3.15 support, a base_url argument for ZaiProvider, and adjustable reasoning for mistral-large-4 models.

The changelog flags a compatibility change: BrowserUse now runs on the host model by default, and its telemetry is opt-in. Fixes also apply auth to requests sent through MCPToolset(http_client=...) without closing a user-supplied client, and propagate Exa 402 payment-required errors instead of retrying them.

If you use BrowserUse, check whether it now routes through the host model and whether you need to opt back into its telemetry. If you pass your own http_client to MCPToolset, auth now applies to those requests and that client is no longer closed by the library.

If you call Exa through Pydantic AI, a 402 payment-required error now surfaces immediately instead of being retried, so your code should handle it.