Hello, we are having issues with gemini-3-flash-preview with grounding, showing an enormous amount of search groundings. We are just starting with Gemini API and really don’t understand why we were billed this much.
The AI Studio dashboard shows about 200 requests made that day with about 100 tool calls to ground search. I don’t know why the billing dashboard shows 37000 calls, but I highly doubt every tool call had about 370 individual searches.
These costs are very high and something we did not expect in our budgeting when starting with the Gemini API. Is there any way to fix this?
Thanks for reporting. This is due to a bug on our side which over-counted search grounding usage. Apologies for this. We’ll be removing / refunding erroneous charges.
I’m on the Gemini Developer API (AI Studio, prepaid), using gemini-3.5-flash with the google_search grounding tool for a batch of short Hebrew fact-check calls. Today my prepaid balance dropped far more than my usage dashboard can justify, and the excess tracks one-to-one with the grounded calls. This looks like the search-grounding billing over-count that was acknowledged and “fixed going forward” in April (threads 139138, 139659), still active in late May.
Expected cost at list price ($1.50 / $9.00 per 1M tokens; $14 / 1k grounded search queries):
tokens: ~$0.8
grounding: 71 × $0.014 ≈ $1.0
total expected: ~$2 (a bit more counting a small enrich stage on 2.5-flash)
Actually charged today: my balance dropped the equivalent of ~$18 (about 5-6x expected). Per the billing cost report, the “Grounding with Google Search” SKU was charged <FILL: grounding $> against only 71 grounding events on the usage side.
Clean control that localizes it to grounding: I ran one stage with no grounding (brief + candidate generation, token-only). It billed exactly as expected from its token count. Every run that used the google_search tool billed ~5-6x its token+grounding cost. So my token rate is correct; the inflation is grounding-only.
Additional detail: across 176 logged responses today (audit dataset export), 0 contained groundingMetadata / a populated webSearchQueries array, even on calls that clearly grounded (they returned cited sources). So there is no client-side way for me to reconcile the billed grounding count.
Questions:
Is my account affected by the search-grounding over-count billing misconfiguration (the one confirmed in threads 139138 / 139659)?
The usage dashboard shows 71 grounding events; can you confirm how many grounded search queries I was actually billed for today, and reconcile the difference?
If this is the same over-count, I’d like the erroneous grounding charges reviewed and refunded, across the whole affected period, not just today (see the multi-day numbers below).
AI Studio’s cost view has no per-SKU breakdown, so a customer cannot isolate or audit grounding charges separately. This both hides the over-count and makes it impossible to self-verify. Can per-SKU cost reporting be exposed in AI Studio?
Happy to share exact request IDs / timestamps. Thanks.
