One support recording can raise several questions. Which team should get it? Is anything urgent? Does a person need to listen?
Google launched the Gemini Decisions API on November 11. Gemini Decision 1 can score yes/no, multiple-choice, and rubric questions against a shared input that includes text, images, audio, video, or PDFs. Google lists a limit of 256K input tokens and up to 1,000 questions per request. The API is available through Google AI Studio, where Google offers a no-cost tier for prototyping.
Google charges $0.10 per million input tokens on the paid tier and lists no output-token charge. The shared material is billed once per call. Question instructions and criteria add their own tokens. Google reports median end-to-end latency of 150 to 250 milliseconds for its tested workloads of up to 16K tokens. The reported near-flat latency comparison covered 1 to 100 questions.
Budget for the decisions and the fallback
Suppose 100,000 calls each contain 10,000 billed input tokens, including the questions. That is one billion input tokens, or $100 at Google’s listed rate. If your test sends 8% of cases to a reviewer or another model, that adds 8,000 follow-up cases. The 8% is an example, and the cost of those follow-ups depends on your workflow.
Take a small set of calls with routes assigned by your support team. Track both incorrect automatic routes and the share sent for review as you change the confidence threshold. Run the same set through any alternative you are considering. Our coverage of OpenAI’s Decisions API gives a starting point for comparing latency and routing quality.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
