What changed
Pydantic AI shipped v2.9.0 on July 11, 2026. The release includes several changes, but the most product-relevant one is a new runtime surface: tools and capabilities can now read the run's usage limits directly from RunContext.
The underlying implementation is spelled out in the merged usage limits PR. Its practical effect is simple: a tool no longer has to rely on duplicated app-level configuration to understand the budget it is operating under. It can compare current usage with the run's effective limits and adjust behavior inside the framework's own runtime.
The same release also pulled in GPT-5.6 model support and reasoning-mode controls, as described in the merged GPT-5.6 reasoning-mode PR. That matters because it means model-selection controls and runtime-budget controls are landing in the same product layer instead of being split across ad hoc wrappers.
Why it matters
This is a more meaningful change than a typical SDK compatibility update. Budget-aware agents usually break down at the boundary between orchestration logic and tool logic: the app knows the limits, but the tool doing the work does not. That often leads teams to re-thread the same limits through custom parameters, duplicate config, or brittle guardrails.
By exposing usage limits in RunContext, Pydantic AI is making budget awareness part of normal agent execution. A code tool, research tool, or capability hook can decide to degrade gracefully, shorten work, skip optional steps, or ask for approval before it burns through the remaining budget. That is a product improvement, not just an internal refactor.
For portfolio and production builders, the broader signal is that agent frameworks are competing on runtime ergonomics now. It is not only about which models they support. It is about whether the framework helps you express practical constraints like cost, rate, and reasoning depth where the agent actually makes decisions.
The takeaway
Pydantic AI is moving one step closer to a framework where budget policy lives inside the agent runtime instead of in side-channel app code. If that direction continues, builders get a cleaner path to agents that are not just capable, but operationally aware.