With store=true and previous_interaction_id on gemini-3.1-pro-preview, a follow-up turn after a 24.7k-token turn reports usage.total_input_tokens = 11, while an undocumented field raw_prompt_token shows 24,750, and the model clearly reads the earlier content. The Interactions docs say usage includes the conversation history.
Questions:
- Is the stored history billed as input on each follow-up turn, and at which rate (standard, cached, or not at all)?
- Which usage field reflects the billed input tokens?
- Does implicit caching apply to the stored history, and is it reported in cached_tokens?
Context: a multi-turn agent with Google Search, URL context and function calling, about 30 follow-up turns per task. We’re deciding whether to move from generateContent to Interactions and need the billing to be predictable.