Google Gemini API Skyrockets in Cost with "Ghost Billing" Bug: Deleting Cache Still Charges, Zero Output Also Billed

According to Beating Monitoring, recently, the Google AI Developer Forum reported multiple urgent pleas regarding the out-of-control billing system of Gemini API. Several developers faced massive unexpected charges during normal usage due to underlying system vulnerabilities, such as someone being charged nearly 27k RMB within just 12 hours. Currently, Google's billing and technical teams are still passing the buck on this matter, with no official fix or quick refund channel announced.

Investigation shows that the main bugs causing developers to receive exorbitant bills are twofold: first, the "Ghost Cache" vulnerability, where context caches created via API expire or are deleted, leaving the front-end management list cleared, but Google's backend continues to bill at thousands of yuan per hour for "idle" processing; second, the "Thinking Dead Loop" trap, where enabling tools like internet search disables the model's "thinking budget limit," causing the model to fall into infinite reasoning when handling simple tasks, exhausting up to 64k tokens before crashing due to timeout. Even when producing "zero output" (no useful response), Google still charges the full amount, with a 1,500-fold increase in reasoning costs.

Due to severe delays of 32 to 72 hours in Google's cloud billing system and the lack of automatic quota throttling mechanisms, developers are being charged large sums before receiving alerts. With official customer service evading responsibility and no direct responses on forums, some affected developers have announced they will completely abandon Gemini's context cache and reasoning models in production environments to mitigate financial risks.