Google’s Gemini Decision 1 puts bounded choices behind a dedicated API

Google has introduced Gemini Decision 1, a model that scores fixed choices instead of generating a prose answer. The Gemini Decisions API accepts one shared input and up to 1,000 yes/no, multiple-choice, or rubric-scoring questions in a request. Google says it can process text, images, audio, video, and PDFs, return probabilities and confidence scores, and handle up to 256K tokens of context. This gives developers a direct way to ask for a classification, tool choice, or policy check without asking a general-purpose model to explain its decision first.

The API became available through the Gemini API in Google AI Studio on November 11. Google lists pricing at $0.10 per million input tokens and zero billed output tokens; it says the shared input is charged once, while each question adds its own instruction and criteria tokens. Google reports 150–250 ms end-to-end latency and nearly flat performance from one to 100 questions per request in its own tests. Those figures are provider claims, not independent measurements, so teams should benchmark their own input sizes and workload before treating them as service-level expectations.

The product sits in a growing category of APIs that return structured decisions rather than generated text. Google’s design combines multiple questions with multimodal input, and the company says one call can mix yes/no checks, choices, and rubric scores. Vercel listed Gemini Decision 1 in its AI Gateway on launch day, offering another integration path. That confirms ecosystem availability, but it does not validate Google’s benchmark comparisons or establish that this model will outperform a text model for any particular task.

For developers, the clearest fit is a bounded branch in an application: route a request, select from a known set of tools, or score an item against a defined rubric. Returning probabilities can make it easier to set a fallback threshold than parsing free-form text, and charging only for input may help when many questions share the same context. Still, confidence scores are not proof that an answer is correct, and Google’s launch benchmarks have not been independently replicated in the sources reviewed here. Treat the endpoint as a candidate for controlled evaluation, especially where a wrong decision could trigger a consequential action.