Google’s Gemini Decisions API trades generated answers for scores

Google introduced the Gemini Decisions API on November 11, pairing a new endpoint with the Gemini Decision 1 model. Instead of generating an answer token by token, it returns probabilities and confidence scores for bounded tasks: yes/no checks, choices among candidates, and ordered rubric scores. The input can combine text, images, audio, video, and PDFs, and one request can contain up to 1,000 questions. That design is aimed at classification and routing work where an application needs a result it can act on, not a paragraph to interpret.

Google says the API accepts inputs up to 256,000 tokens and charges $0.10 per million input tokens, with no output-token fee. It also reports 150 to 250 milliseconds end-to-end latency and says shared input is billed once regardless of question count, while each question adds instruction and criteria tokens. An independent Japanese technology report confirms the endpoint’s launch, its multimodal input, the three decision formats, and the stated price. The detailed timing, billing behavior, and performance figures remain Google’s own claims; the independent report does not provide a separate benchmark.

The practical comparison is with a conventional generative call that reads an input and writes a response which downstream code must parse or judge. Google says Gemini Decision 1 scores options in a single forward pass and presents a comparison in which it is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API across a shared benchmark suite. That is a provider-run comparison, not evidence that it will be five times cheaper for every application: question count, prompt length, modality, existing classifier costs, and fallback rates all affect the result. The API’s bounded outputs may simplify integration, but they also require developers to define suitable candidate choices or scoring rubrics up front.

For developers, the useful test is whether bundling many decisions over one shared input reduces total latency and cost without hurting accuracy on the cases that matter. Confidence scores can support a fallback path, but they are not a guarantee that a classification is correct or calibrated for a team’s data. Google describes internal use cases such as model routing and tool safety checks, and says integrations with several third-party frameworks are rolling out. Those examples and availability details are from Google; the independent coverage confirms the launch basics, not production performance. Teams should measure error rates, tail latency, and the cost of ambiguous cases before replacing a generative call or a specialist classifier.