Clarity for AI buildersSponsored by AI Routing
availability

Google introduces Gemini Decisions API for up to 1,000 closed-ended questions per pass

Google introduces Gemini Decisions API for up to 1,000 closed-ended questions per pass
Google's Gemini Decisions API handles up to 1,000 closed-ended questions per pass, according to The Decoder. The Decoder reports that Google DeepMind is introducing the API with an underlying model called Gemini Decision 1. It supports yes/no questions, multiple-choice selections and scale-based ratings across text, images, audio, video and PDFs. Rather than generating an open-ended response, it scores options directly and returns probabilities with confidence values. The report cites response times of 150 to 250 milliseconds. That makes the reported offering more narrowly focused than a general-purpose conversational interface: applications supply defined questions or choices, and the model evaluates them. The capacity claim concerns up to 1,000 independent questions in one pass, not a sequence of 1,000 dependent reasoning steps. It should therefore not be read as evidence of equivalent performance on a long, multi-step workflow. According to the report, Google says it already uses the API internally for dynamic model routing in Gemini Notebook, design prototyping in Stitch and safety checks in Google Search. A Google example described in the article evaluates a kitchen-counter photo using three question types: stain detection with confidence of 0.95, selection of a cleaning action at 98 percent, and a severity rating of 1.19 out of 2. These illustrate supported output formats, rather than establish accuracy across other images or tasks. Pricing is $0.10 per million input tokens, with no output-token charge, The Decoder reports. Google describes that as about five times cheaper than Gemini 3.5 Flash-Lite. The API is reportedly available through Google AI Studio, while integrations with OpenRouter, LangChain, Vercel and other platforms are rolling out. Those integrations should not be treated as universally available on the strength of this report. For developers, the practical distinction is the combination of predefined choices, confidence information and input-based pricing. This could make the API relevant to routing or classification workloads resembling Google's cited uses. However, the supplied report does not specify the conditions behind the latency figures or provide enough benchmark methodology to establish how performance would transfer to a particular deployment. The article also describes a Google benchmark graphic showing Gemini Decision 1 leading categories including agent routing, reasoning and multimodal understanding, with the lowest Top-1 expected calibration error, or ECE. That calibration claim concerns the relationship between confidence and correctness; it is not a guarantee that individual decisions will be right. The supplied evidence does not independently verify Google's performance or cost comparisons.