Google launches Gemini Decisions API for fast multimodal choices

Google introduced the Gemini Decisions API on 11 November, giving developers a new way to build software that makes defined choices from documents and media. Available through the Gemini API in Google AI Studio, it uses the new Gemini Decision 1 model to return probabilities and confidence scores.

A request can evaluate up to 1,000 independent questions against shared text, images, audio, video or PDFs. Developers can combine yes-or-no checks, selections from a fixed list and ratings against an ordered rubric. Google's model documentation lists a context limit of 262,144 tokens, covering the shared material and all questions. For a support application, that could mean checking a customer's message and photograph together before deciding which team should handle the case.

Google lists a paid price of $0.10 per million input tokens, with no output tokens generated or billed. A no-cost tier is also available. Shared material is billed once within each request; instructions and criteria for individual questions add to the input count. Sending the same material in a separate request incurs input charges again, and context caching is unsupported.

Speed is a central part of the announcement. Google reports median end-to-end latency of 150–250 milliseconds in tests using contexts up to 16,000 tokens, with latency staying nearly flat between one and 100 questions. Those are the company's measurements under stated conditions, rather than a guarantee for every workload or the maximum question count.

Vercel separately announced that Gemini Decision 1 is available through its AI Gateway. Its integration supports decision calls alongside other model requests, with shared usage tracking and budgets. Vercel specifies AI SDK 7 for audio, video and PDF inputs; its TypeSafe-compatible and Decisions interfaces support text and images.

For businesses and makers, the practical opportunity is to put small, repeatable judgments inside larger workflows: routing requests, screening content or selecting an agent's tools. Reliability still depends on how those choices are defined and tested. Google's developer guide recommends setting confidence thresholds using labelled examples from actual traffic and considering the cost of different mistakes. It also warns that a multiple-choice question distributes probabilities across the supplied options even when none fits. A fallback category or human review therefore remains a meaningful design decision.