The release of Pydantic AI v2.64.0 on October 30, 2026, introduces an opt-in sticky setting for ModelFallback, as detailed in the official release notes and the associated pull request #11330. This feature allows a conversation to remain on a fallback model that has successfully answered a request, rather than defaulting back to the primary model at the start of every subsequent request step. By enabling sticky=True, builders can alter the default model selection logic to respect the recent failure history of a specific conversation. This configuration is disabled by default, ensuring that existing agent behavior remains unchanged unless explicitly opted into.
The scope of this stickiness is strictly limited to an individual conversation. When enabled, the selected model persists through message_history, even when that history is passed across different processes or handled via durable execution. It is important to distinguish this from a global circuit breaker. The setting does not track the health of a model provider across the entire application or for other agents. Instead, it modifies the starting point for the next step of a specific conversation. Because it operates at the conversation level, it does not replace the need for explicit model selection or establish a platform-wide health state for any given model provider.
The rationale for this implementation centers on the behavior of primary models during periods of downtime or rate limiting. Without sticky=True, each step whose primary is down or rate-limited waits for the primary to fail, spends retries, and switches providers back and forth. The pull request notes that these repeated switches can result in the loss of prefix caching. By pinning the conversation to a successful fallback model, the system is designed to avoid the overhead of repeatedly attempting a primary model that has already demonstrated failure for that specific conversation path.
The mechanism includes a defined wraparound logic to handle potential failures of the sticky model itself. If the sticky model fails, the system attempts to use the models listed after it in the fallback chain, followed by the step’s own model, and finally any models that appeared before the sticky model in the sequence. Each model in this chain is attempted at most once. The primary model may be included in this sequence, and if it successfully answers, the conversation will return to it. This logic governs how the system attempts to recover when the sticky model is no longer available.
A critical aspect of this design is the absence of an expiry timer. The system does not automatically recheck the primary model after a set interval. Recovery is not guaranteed; it is entirely dependent on the fallback chain successfully producing an answer or the primary model being reached during the wraparound process and successfully responding. This creates a specific tradeoff for builders. While the feature provides continuity in routing after a primary failure, it places the responsibility on the developer to manage the conversation state, as no timer-based recheck is provided to force a return to the primary model.
Existing model selection mechanisms maintain their priority over this new setting. The sticky behavior only affects the agent’s default model selection. Explicit instructions, such as those provided via run(model=…), override(model=…), or SelectModel, continue to take precedence. This allows builders to implement sticky failover paths for default operations without overriding deliberate model choices made for specific tasks or complex agent workflows.
This update provides a narrow, configurable tool for managing conversation routing. It alters the retry and provider-switching path to favor the last successful fallback model. The release documentation outlines the mechanism and its constraints, but it does not offer data regarding production performance, latency, or cost impacts. Builders should evaluate this feature based on the specific requirements of their conversation state management.