I’m looking for clarification on some of the terms and conditions for the Gemini api, not sure where to find any information. Any help is appreciated.
Use case explenation:
We use the Gemini Developer API (GEMINI_API_KEY, the @google/genai SDK), not Vertex. Our pipeline makes two calls per report — call 1 is grounded with Google Search and produces a briefing with groundingChunks and groundingSupports; call 2 takes that briefing as input, is not grounded, and restructures it into the schema we display. What the end user reads is call 2’s output, carrying citations derived from call 1’s metadata.
**
Questions:**
A. Architecture decisions
1. Is the output of call 2 still “Grounded Results” for the purposes of the Additional Terms? Call 2 makes no search of its own; it reshapes call 1’s grounded briefing. We need to know whether the display, caching and single-user restrictions attach to it. Why it matters: every other question below inherits this answer. A yes makes our retention model non-compliant; a no makes most of this moot.
2. Does the prohibition on caching Grounded Results prevent serving one grounded generation to more than one end user? We generate a report once and serve it to multiple users who search the same company within a retention window (measured ~1.74 users per generation). The terms say display is limited to “the end user who submitted the prompt”, with a two-year exception for storage in an end user’s own history. Ask specifically: is per-user re-grounding required, or is a shared generation permissible if each user’s view carries the required Search Suggestions? Why it matters: this is a ~19% cost swing for us, not a formatting change. It decides whether the product’s unit economics work as designed.
B. Billing
3. When a grounded call returns ZERO groundingChunks, are the underlying search queries still billed? We retry when a response comes back ungrounded, and discard the first attempt. Ask specifically: is a discarded ungrounded attempt billed at the per-query grounding rate, and if so, is the query count recoverable from the response for reconciliation? Why it matters: at our query volume one discarded attempt is worth roughly 8x the entire token cost of a report. If ungrounded attempts are not billed, our cost model is materially wrong in our favour; if they are, this is our single largest variable cost line.
4. How is the 5,000 free grounded prompts/month allowance scoped and counted? Per project, per API key, or per billing account? Does a call that returns zero groundingChunks consume one? Does a call issuing multiple search queries consume one prompt or several? Why it matters: we model the free tier as covering a specific monthly report volume, and every part of that arithmetic depends on the counting rule.
5. Is toolUsePromptTokenCount billed, and at which rate — input or output? We have observed it being large while our receipt recorded almost nothing. Why it matters: we have just wired it into our ledger and need to price it correctly rather than guess.