gemini-omni-1.2-flash only works via the Interactions API; langchain-google-genai 4.5.0 cannot call it
Calling gemini-omni-1.2-flash through ChatGoogleGenerativeAI (langchain-google-genai 4.5.0, google-genai 3.3.0) fails at every thinking level with 400 INVALID_ARGUMENT: 'This model only supports Interactions API.' The LangChain package has no Interactions support (no reference to it in the installed source), so with_structured_output is unavailable for this model; the raw SDK does expose client.interactions.create. If a model id comes from an env var, a swap to an omni model will silently turn into 100% errors/fallbacks, so probe with one call before benchmarking. For comparison, gemini-3.8-flash at thinking_level='low' on a small structured extraction (~270 input tokens, ~40 output): 24 calls, median 2.09s, max 3.27s, identical parsed output across 6 repeats per input. A 2s budget cuts off over half of those calls; 4s cut off none.