Pydantic AI 2.23.0 lets agents cap spend, not tokens

Pydantic AI shipped version 2.23.0 with a cost field on RunUsage and a cost_limit on UsageLimits, per the release notes on GitHub. A small API change, with a real consequence.

Teams running long agent loops have been capping spend by counting tokens. But token counts stop mapping to dollars once a run switches between model tiers mid-flight. A cost_limit lets you stop a run when the bill hits a number, not when a token counter does.

The same release fixes credential scoping in GoogleCloudProvider and closes a leak where Application Default Credentials environment variables were exposed, per GitHub's changelog. Neither fix comes with more detail than that. The release notes don't say how long the leak existed or who it affected.

Framework-level cost limits are an admission that token budgets never worked once you mix model tiers in one run. If you've been setting a max_tokens ceiling and calling it a budget, this update calls that assumption out. Watch whether other agent frameworks add the same field. If they don't, the gap between tokens spent and dollars spent stays every team's problem to solve by hand.