Google releases Gemini Decisions API to cut latency and costs on bounded choices

Google launched the Gemini Decisions API on November 11, 2026, introducing a service designed to answer targeted, closed-set questions at lower costs than traditional generative models. Available via Google AI Studio, the tool runs on a specialized Gemini Decision 1 model and evaluates up to 1,000 questions in a single pass across text, audio, images, video, and PDFs. Google charges $0.10 per 1M input tokens with zero output tokens billed, citing an end-to-end response latency of 150-250 ms. [1][2]

Rather than producing text token by token, the model scores candidate options during a single forward pass, delivering confidence ratings from 0 to 1 alongside normalized probabilities. It supports multimodal inputs up to 256K tokens and handles three specific formats: yes-or-no evaluations, multiple-choice selections between 2 and 255 items, and ordered rubric grading. Google reports that median model compute latency stays between 50-100 ms for contexts up to 16K tokens, remaining nearly flat whether assessing one question or 100. [2]

The API is aimed at operational workloads such as query routing, policy verification, content grading, and tool selection. Google says internal teams already use it as a control plane; Gemini Notebook employs the system to route prompts to different model tiers, which the company claims reduces everyday query latency by up to 40%. Compared across a shared benchmark suite, Google claims the service operates nearly 5x cheaper than running Gemini 3.5 Flash-Lite through its Interactions API. [2]