What changes compared with generative models
Google DeepMind presents the Gemini Decisions API, an interface designed for rapid closed-ended decisions: yes/no questions, multiple choice, and scale-based ratings. The underlying model, Gemini Decision 1, processes up to 1,000 independent questions in a single pass and across multiple formats—text, images, audio, video, and PDF. According to the source, the system assigns a direct score to the options and returns probabilities with confidence values within a range of 150 to 250 milliseconds.
The new element is therefore not a more powerful conversational model, but a model specialized in classifying and choosing. It is a difference that matters to those building systems that must make many elementary decisions quickly, where text generation would be superfluous and costly.
Stated costs and internal uses
The company states a price of $0.10 per million input tokens, with no charge for output tokens, and claims the cost is about five times lower than that of Gemini 3.5 Flash-Lite. These are figures provided by Google and not independently verified on the basis of the available material.
Google says it already uses the API internally in three contexts: dynamic model routing in Gemini Notebook, design prototyping in Stitch, and safety checks in Google Search. The API is available immediately through Google AI Studio, while integrations with OpenRouter, LangChain, Vercel, and other platforms are described as being rolled out. No timeline is given for completing these integrations.
A crowded field, with issues still unresolved
The launch fits into a recent trend: according to the source, after the startup Jev came similar offerings from OpenAI, Cloudflare, and Microsoft. Competition is therefore shifting from generation to fast, low-cost evaluation, a segment of interest for applications that need many repeatable decisions rather than elaborate answers.
However, aspects remain to be clarified. The material does not specify how the 150–250 milliseconds are measured or under what load, and it does not clarify the reliability of the probabilities returned for ambiguous questions or across different formats such as audio and video. Nor does it indicate whether the 1,000 queries per pass have limits on length or complexity. It is precisely these details that determine whether the model will be suitable for decisions with operational consequences, such as the safety checks cited by Google itself.
