Users of the Google Gemini API report sky-high billing errors

According to monitoring by Beating, recent reports on Google’s AI developer forum have revealed multiple urgent appeals regarding the uncontrolled billing system of the Gemini API. Several developers, while using the service normally, have faced massive unexpected charges due to underlying system vulnerabilities—for example, one developer was charged nearly RMB 27,000 within just 12 hours. To date, Google’s billing and engineering teams continue to deflect responsibility, with no official statement or expedited refund process issued. Investigations have identified two core bugs responsible for the exorbitant bills: First, the “phantom cache” vulnerability—after context caches created via the API expire or are deleted, and the frontend management interface shows them as cleared, Google’s backend billing system continues charging at rates of thousands of yuan per hour as if the cache were still active. Second, the “infinite reasoning trap”—when tools like web search are enabled, the model’s “thinking budget limit” fails, causing it to enter an infinite loop during simple tasks, consuming up to 64,000 tokens before timing out and crashing; even when the final output is “zero” (no useful response returned), Google still charges full price for thinking, with fees spiking up to 1,500 times normal levels. Due to severe delays of 32 to 72 hours in Google Cloud’s billing system and the absence of automatic spending caps or circuit breakers, developers are hit with massive charges before receiving any alerts. With official customer support deflecting blame and no substantive responses on forums, some affected developers have announced they are completely disabling Gemini’s context caching and reasoning models in production environments to mitigate financial risk.