Newer gemini models support things like grounded searches and timed caching. However openai compatible api might not expose these costs in the input/output tokens the way claude api does.

verify that metering for this is valid, and fix if not