Menu

27,262 thinking tokens, zero answer—and a “resolved” Google Cloud billing case

Reading 0:00
15s threshold

I can live with an LLM giving me a bad answer. What I can't accept is an API throwing a native generation error, returning zero visible text or tool calls, burning 27,262 internal thinking tokens in the process, billing me for it, and Google Cloud Support closing the ticket as "Resolved" without giving me a single line-item audit.

Before anyone asks: this is on an upgraded Paid billing account, though the charges were absorbed by my remaining promotional credits rather than my credit card. That doesn't make this a non-issue. Promotional credits are finite, and the billing rule Google just confirmed in writing applies to every paying developer running Vertex AI in production.

The Receipt: One Single Call on Oct 29

Here is one of the failures from my Vertex AI logs (model: gemini-4-argon ):

Timestamp: 2026-10-29 07:17:52 UTC

Google Response ID: aPHiao_zIaezi9sPqNjBuQc

Billed Usage: 2,049 uncached input | 531,163 cached input | 27,278 output tokens (27,262 of which were internal thinking tokens)

Native…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.