Description

Publishing here as a continuation of the conversation on Slack.

Currently, FallbackModel only accepts a list of model instances (or names) and a fallback_on predicate, but there’s no way to supply different ModelSettings (e.g. thinking_config) per model.

That means if you want to use, say, Gemini 2.0 Flash (which doesn’t support a thinking budget) and Gemini 2.5 Flash (which does), you cannot configure each individually—you can only pass a single model_settings at the Agent or run call level (which applies to all fallback models).

This leads to failures when the first model in the fallback chain doesn’t support settings required by the second (and vice versa).

References

FallbackModel