What changed
Pydantic AI shipped v2.23.0 on August 4, 2026 with two product-level additions that matter more than a normal framework patch. First, it added run-level cost accounting and caps through RunUsage.cost and UsageLimits.cost_limit. Second, it added ToolAvailabilityDeltaPart, a first-class way to represent tools becoming available mid-run instead of faking a model-driven tool search in history.
On August 5, 2026, v2.24.0 followed immediately with hardening around the same surface area: preserving caller-owned OpenAI model settings, keeping stream part indexes stable during replay, and tightening tool-approval typing for Vercel AI integrations.
Why it matters
A lot of agent frameworks expose token counts but still leave actual spend controls and live tool-surface changes to application glue code. Pydantic AI is moving both concerns into the runtime. That means builders can set budget limits in dollars, inspect cost as part of normal run usage, and tell the UI exactly when an agent gains a new capability without inventing synthetic history events.
That combination is useful for real product work: cost limits are easier to turn into plan tiers, admin controls, or user-visible safeguards, while native tool-availability deltas make approval-driven or deferred-tool interfaces easier to explain in the product itself.
Product takeaway
The practical signal here is not just that Pydantic AI added more features. It is that the framework is getting more opinionated about production concerns that teams usually bolt on later: budget enforcement, portable history for changing tool access, and quick follow-up fixes across OpenAI, Google, Bedrock, and Vercel-facing integrations. For teams building agents into real apps instead of demos, that is a meaningful shift.