Google’s current token documentation states that usage.total_input_tokens on a previous_interaction_id continuation includes conversation history, and that cached history is represented in total_cached_tokens. We understand these as documented post-dispatch usage reporting. We are not asking Google to reconfirm those facts. Reference: Understand and count tokens.
Is there a supported pre-dispatch method to count or finitely upper-bound the exact native Interactions request, including full explicit history, opaque thought signatures, system instruction, custom-function declarations/results and any protocol overhead, before the request is submitted?
By ‘upper bound,’ we mean a request-specific bound derived from the actual supplied request that is materially tighter than simply reserving the model’s entire maximum context window.
For a complete explicit-history request (store:false, without previous_interaction_id), please identify the supported method/endpoint and request schema, applicable model/version constraints, and the scope of its guarantee. Does the result bound all chargeable input tokens for that exact request, including any cached and uncached input and overhead? If something is excluded, please identify the documented finite bound for that exclusion.
The CountTokens reference documents contents and generateContentRequest; we have not identified a documented schema or guaranteed mapping covering the complete native Interactions input described above.
If no such supported pre-dispatch method or guaranteed finite bound is currently available, please state that clearly and identify Google’s recommended mechanism for applications that must enforce a hard monetary ceiling before dispatch. Please distinguish enforceable guarantees from estimates, and explain any required constraints or limitations. Post-dispatch usage receipts alone cannot make an advance admission decision.