Google Gemini API Sky-High "Ghost Billing" Vulnerability: Deleting Cache Still Charges, Zero Output Also Billed

According to “Beating Monitoring,” recently, multiple emergency requests for help regarding the Gemini API billing system going out of control were reported on the Google AI Developers Forum. During normal use, multiple developers faced massive abnormal deductions due to vulnerabilities in the underlying system. For example, one person was charged nearly 27,000 yuan RMB in just 12 hours. At present, Google’s billing team and technical team are still trading blame for this matter, and no official fix statement or quick refund channel has been released.

After investigation, the two main core bugs that led to developers receiving sky-high bills are as follows: First is the “ghost cache” vulnerability. When the context cache created via the API expires or is deleted, the front-end management list is emptied, but Google’s backend billing continues to “spin idle,” deducting fees at a rate of thousands of yuan per hour. Second is the “thinking dead loop” trap. When tools such as enabled network search are turned on, the model’s “thinking budget limit” fails, causing the model to get stuck in infinite reasoning while handling simple tasks. After burning through up to 64,000 tokens, it times out and crashes. Even if the final result is “zero output” (no useful answers are returned), Google still charges the full amount, and the thinking cost skyrockets by 1500 times.

Because Google Cloud’s billing system has a severe delay of 32 to 72 hours and lacks a quota-based automatic circuit-breaker mechanism, developers had already been charged large sums before they received any alerts. Due to official customer service deflecting responsibility and no one giving a direct response in the forum, some affected developers have announced that, to avoid financial risk, they will completely abandon Gemini’s context cache and reasoning models in production environments.