Google launches an AI decision API that scores up to 1,000 questions per request

Google is giving developers a way to ask AI for a decision without waiting for a written reply. Launched November 11, the Gemini Decisions API evaluates up to 1,000 questions against one shared input. Its new Gemini Decision 1 model targets jobs such as choosing tools, checking policies and routing requests—not composing answers.

Supply the choices before asking the question

The central change is how the model produces its result. Instead of writing an answer one token at a time, Gemini Decision 1 scores candidate options in a single processing pass. Developers define the questions and allowed answers; the service returns probabilities and confidence scores. That makes this an interface for bounded choices, rather than an open-ended conversation.

A shared input can combine text, images, audio, video and PDFs. Google gives examples such as pairing a call recording with account information, or a scanned insurance claim with damage photos. Multiple questions can examine that same material in one request, mixing three kinds of decision rather than requiring a separate call for each format.

• Yes/no checks estimate the probability that a condition holds. Developers can optionally describe what should count as true or false.
• Multiple-choice questions accept between two and 255 options. The API returns the winner and probabilities for every candidate, not just the selected answer.
• Rubric scoring uses an ordered scale starting at zero. It returns a continuous weighted score alongside probabilities for each level on that scale.

Google’s sample combines a breakroom photo with a note about an upcoming meeting. One request checks for a spill, chooses a cleanup action and rates the mess. Each answer also includes a confidence score between zero and one. Google says applications can use those scores to send ambiguous cases to another system or a human reviewer.

One shared input, one input bill

The billing structure follows that shared-input design. Gemini Decision 1 costs $0.10 per million input tokens, the units used to measure material sent to the model. The shared material is billed once per request, regardless of the question count. Each additional question adds only its own instructions and criteria to the input bill.

There are no output-token charges, variable thinking-token bills or separate charges for audio transcription and reading text from images. Developers can access the service through the Gemini API in Google AI Studio, where Google also offers a no-cost tier for prototyping. Its launch includes Python, JavaScript and REST guides.

256K tokensMaximum input

Google lists support for inputs up to 256K tokens.

Up to 16K tokensContext covered by speed figures

Google’s latency description covers contexts up to 16K tokens, not the full input limit.

The speed claim has a defined envelope

Google reports 150–250 milliseconds of end-to-end latency, including 50–100 milliseconds of median model processing time. Its benchmark description covers multimodal inputs and says latency remains nearly flat between one and 100 questions per request. Those figures are narrower than the service’s maximum capacities: they are not a speed promise for every supported input size or all 1,000 questions.

The company also says Decision 1 has the lowest calibration error across its benchmark suite, spanning routing, reasoning, ranking and multimodal understanding. Calibration concerns how closely a model’s confidence matches its actual correctness. Google reports that the Decisions API is nearly five times cheaper across the shared suite than Gemini 3.5 Flash-Lite accessed through its Interactions API.

One internal use shows where Google expects the mechanism to pay off. Gemini Notebook uses the API to choose a model tier and thinking level for incoming requests. Google says that routing cuts latency for everyday queries by up to 40%, while reserving deeper reasoning models for complex tasks. Here, the decision model’s job is to choose which other model should do the answering.