Gemini Decisions API: 1,000 Questions per Call, No Output Fees

Google’s Gemini Decision 1 charges $0.10 per million input tokens and nothing for output, and it answers up to 1,000 yes/no, multiple-choice or rubric questions against one shared input in a single pass. For agent builders who still spend a generative call on every routing choice or policy check, the bigger change is the input: it is the first of the decision models Simon Willison compares that accepts audio, video and PDFs.

What Happened

Google announced the Gemini Decisions API as an endpoint for “fast, closed-set decisions” powered by the gemini-decision-1 model, available in Google AI Studio. A request carries one multimodal input of up to 256K tokens and up to 1,000 independent questions. Instead of generating answer text token by token, the model scores the candidate options in one forward pass and returns probabilities and confidence scores, with 150 to 250 ms end-to-end latency according to Google.

There are three question types: noul for yes/no, choice for 2 to 255 options, and score for an ordered rubric. Google says it already runs the API internally for model routing in Gemini Notebook, where it serves everyday queries with up to 40% lower latency, and as a safety layer for MCP tools in Search. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, with Gemini in Google Cloud listed as coming soon.

Why It Matters

The billing model rewards asking many small questions. Google says the shared input is billed once per request no matter how many questions are evaluated, and each question adds only its own instruction and criteria tokens. There are no output-token, thinking-token, transcription or OCR charges. On Google’s own benchmark suite, it says Gemini Decision 1 is nearly 5x cheaper than Gemini 3.5 Flash-Lite through the Interactions API.

Willison’s write-up places the launch in a category that now has three hosted options. Gemini’s price matches OpenAI’s GPT-6 Luna decision model at 10 cents per million input tokens, while Jev “remains the cheapest of the three, at 4.2 cents per million.” The difference is input type: he notes Jev is text and JSON only, OpenAI handles text and images, and Gemini covers text, images, audio, video and PDFs. He also points out that the API is “almost identical to Jev” and even reuses Jev’s term “Noul.”

One behavior differs in practice. In Willison’s video test, Jev and OpenAI return 1.0 or 0.0 for yes/no questions, but Gemini returned values such as 0.92 for “robot.” The Decisions API documentation confirms the design: yes/no probabilities always fall between 0.01 and 0.99, so “a check above 0.99 never matches.”

What This Means for Builders

The documentation pushes a specific style: ask narrow, single-purpose questions and combine the answers in ordinary code, using the confidence value to decide whether to act automatically or escalate to a fallback or a person. Teams porting thresholds from another decision model should re-tune them, since a rule written for a 1.0 answer will never fire. Choice options are sorted alphabetically by key before scoring, and exact ties go to the first key.

Stack Archive’s coverage of OpenAI’s Decisions API and the System One API standard tracked the same free-output pricing. Google’s version extends that pricing to multimodal inputs, such as a call recording paired with account data. Google’s claim that it “leads across the suite, with the lowest calibration error” is its own benchmark, not an independent one.

Source: Simon Willison's Weblog

Content created with AI assistance and reviewed for accuracy.