Prepaid credits depleted ~10x faster than expected — gemini-flash-latest silently repointed to 3.6 Flash with thinking-on default; requesting goodwill adjustment Body (fill in the brackets):

Gemini API
1

Billing account ID: 01CB7F-19F1B2-1905..
Period: July 21–29, 2026

Our research tools call the Gemini API for short, JSON-only classification prompts
(temperature=0, expected ~50 output tokens per call), using the gemini-flash-latest
alias. Around July 21 this alias began resolving to Gemini 3.6 Flash, which (a) has
roughly 5x higher token prices ($1.50/$7.50 per M) and (b) enables thinking by default,
with thinking tokens billed at the output rate. As a result, even “answer with JSON
only” calls silently consumed hundreds to thousands of hidden output tokens each.

Impact: ~47,800 short calls consumed approximately $220–250 of prepaid credits.
At the flash-tier prices previously in effect, the same workload was budgeted at ~$21.
Nothing changed on our side; the cost increase came entirely from the silent alias
repointing plus the thinking-on default.

We have since pinned explicit model versions (gemini-2.5-flash-lite) and set
thinking_budget=0 on all calls.

Request: a goodwill credit adjustment for the excess consumption caused by the silent
repointing. Happy to provide usage screenshots from the AI Studio usage page.

2

Hello @Murat_Gurel ,

Note: For all refund related requests, please contact Cloud Billing support

Having said that, gemini-flash-latest “silently” pointing to a new model is an intended behaviour as mentioned in the doc here. As this is intended behaviour, it does not qualify for a refund per se

Also, Default thinking is not newly introduced in 3.6 flash, it was introduced across the entire Gemini 3 series (including Gemini 3 Flash Preview and Gemini 3.5 Flash) prior to 3.6 Flash.

3

Hi @Murat_Gurel ,

Aakash here from Google,
+1 to the answer posted by @Sai_Deepika_K .

Could you also share some example prompts? 3.6 is significantly more token efficient than 3.5 so curious to see what specific prompts are leading to the increased cost. Thanks

4

Hi Aakash, thanks — I’ll take the refund request to Cloud Billing support as advised.

To correct the record on versions: our workload was not migrated from 3.5 to 3.6. gemini-flash-latest had been resolving to 2.5 Flash for us; on ~21 July it jumped a full generation to 3.6 Flash. So “3.6 is more token efficient than 3.5” doesn’t address our case — our unit economics changed from a non-thinking-default tier at roughly $0.10–0.30/MTok to $1.50/$7.50 with thinking on by default.

Example prompt (translated from Turkish; this ran ~46,000 times over a week in a batch library-classification pipeline):

  1. junk: is this NOT a library work (invoice/form/barcode/photo/blank/fragment/ad)? real article/book/thesis = false.
  2. if junk=false, law: is it LAW (economics/finance-technical/sociology=false; legal doctrine/legislation/case law=true)?
  3. if law=true, pick ONE category from: {category_list}
  4. author(surname) + title + year.
    Filename: “{oldname}”
    TEXT: {first_5500_chars}
    Return JSON ONLY: {“junk”:false,“law”:true,“kategori”:“…”,“yazar_soyadi”:“…”,“eser_adi”:“…”,“yil”:“…”}
    Expected output is ~50–80 tokens of JSON. The prompts didn’t get less efficient — the cost driver was hidden reasoning tokens billed at the output rate on a task that needs none (deterministic classification). Two measured data points from our side: (a) on 2.5 Flash with default settings the same class of prompt burned ~160 thinking tokens per call, zero with thinkingBudget=0; (b) at the time of the incident, 3.6 Flash rejected thinkingBudget=0 with 400 INVALID_ARGUMENT, so the reasoning cost could not be disabled on the model the alias had moved us to.

I understand alias drift is documented behavior. As product feedback: when an alias crosses a model generation with a ~10× price change and a billing-relevant default (thinking) that cannot be turned off, existing alias users get a silent 10–50× cost multiplier on unchanged workloads. A notification, a grace window, or generation-scoped aliases would prevent this class of surprise.

Thanks again for engaging on the technical side — the version clarification above should also help anyone else who lands on this thread from a -latest alias surprise. I’m moving the billing side of this to Cloud Billing support and will report back here for the record.

Best regards,
Murat