A serious API billing anomaly has been exposed on Google AI developer forum. A developer reported that the Context Caching feature of Gemini 3 Flash continued to charge at over 1,000 Brazilian Reais per hour even after he had deleted all caches via the frontend API. Within days, the accumulated bill reached 17,847.21 BRL (around several thousand USD), approaching 20,000 BRL.
Charges Spiked Exponentially on June 6
According to BigQuery data provided by developer Danilo_Oliveira, the anomaly began on June 3, 2026. Initially, the "cached text storage token hours" fee stayed at 20–30 BRL per hour with about 4 million token hours. On June 6, usage skyrocketed to over 200 million token hours in a single hour, with hourly charges exceeding 1,000 BRL. By early June 7, total billing was 17,847.21 BRL across 341 charge events.
API Disabled to Stop Bleeding; No Official Fix Yet
The developer immediately shut down the caching script and confirmed via Google's REST API that the frontend cache list was completely empty. Yet the backend continued charging without control. Suspecting a server-side bug where cache records were not properly cleared, he opened billing ticket #720261 and ultimately disabled the entire Gemini API service in his Google Cloud project to prevent further financial loss.
Community Alarm: Cache Feature Turns Into Money Pit
The incident quickly stirred concern among developers. Context Caching was designed to reduce cost and latency for long-text LLM processing, but now it has become a financial sinkhole. Until Google releases a fix, the community advises all developers using Gemini API caching to monitor real-time billing closely and set strict budget caps and alerts to avoid surprise bills.