Google suggests a lightweight classifier to route Gemini models and thinking levels

On November 10, 2026, at 16:10, Google Developers Blog published a post about inference cost management for AI agents. The post argues that as agent loops, tool calls, retries and internal thinking passes pile up, per-token API pricing no longer determines production economics. The metric that actually governs cost, it says, is cost per successful task: total spend across every loop iteration, tool call, retry and internal thinking pass, divided by the number of tasks that pass verification.

The piece notes that when an agent calls a high-thinking reasoning configuration for every micro-decision in its harness, it pays deep-reasoning prices for questions a lightweight configuration already answers accurately. To prevent this unnecessary overhead, the post suggests adding a lightweight routing layer ahead of expensive reasoning calls, using a structured-output classifier from TypeSafe, called Jev, to intercept routine decisions and route only hard problems to high-thinking tiers. Google states that it has no partnership with TypeSafe and that the post is not a product recommendation for the company or its tools. That is where the available coverage ends, with no further developments reported.