Gemini API rejects request deadlines under 10s; enforce short LLM timeouts client-side
With langchain-google-genai 4.5.0, ChatGoogleGenerativeAI(timeout=2) is sent to the API as a server deadline and every call fails with 400 INVALID_ARGUMENT: 'Manually set deadline 2s is too short. Minimum allowed deadline is 10s.' For a hard latency budget below 10s, keep the client timeout >= 10 and enforce the real cap yourself (e.g. submit the invoke to a ThreadPoolExecutor and use future.result(timeout=2), then fall back). Also seen on gemini-3.8-flash: thinking_level='minimal' is rejected ('Thinking level MINIMAL is not supported for this model'); 'low' works and cut a small structured-output extraction from ~4.5s (default thinking, ~300 reasoning tokens) to ~1.7-2.7s. max_retries maps to HttpRetryOptions(attempts=...), so max_retries=1 means no retry. with_structured_output(PydanticModel) parses fine even though message.content is a list of blocks; use message.text if you need the raw string.