Predefined choices instead of text with the Gemini Decisions API

Google’s API evaluates up to 1,000 questions against shared multimodal input. It returns probabilities, with no output-token charges.

Up to 1,000 independent questions can be evaluated against the same document, image, or recording, without generating a written response for each one. Available through Google AI Studio, the Gemini Decisions API uses the new Gemini Decision 1 model to assign probabilities to predefined options. Inputs can include text, images, audio, video, and PDFs.

This structured output can route a request to a model, select a tool, or check content against specific criteria. The program calling the API then uses the scores to trigger an action, consult another AI model, or send the case to a human reviewer. Making a choice and executing it remain separate operations. Checking probabilities against actual results

Each answer includes a confidence score. To be well calibrated, a probability should match the observed success rate across many examples: among answers assigned 80% confidence, roughly eight out of ten should be correct. That number helps set thresholds for automated decisions, but it does not guarantee the outcome of an individual case.

In its comparisons, the provider reports the lowest calibration error among the four systems tested. Its table places the model first on five of the seven shared evaluations, while Jev 1.14 retains the lead on Mind2Web Multimodal and When2Call. The additional audio and video tests compare only two models.

The Artificial Analysis Decision Index offers another perspective: the model’s score of 66 matches the median of the models listed when the page was consulted. This independent index combines several evaluations and accounts for accuracy, confidence, and decision consistency. A score of 66 does not mean 66% of answers were correct, and it is not the same metric Google reports. For deployment, the documentation recommends setting confidence thresholds using examples representative of actual traffic. Three formats, with no dependencies between questions

The API documentation distinguishes between binary questions, multiple-choice questions, and scores based on an ordered rubric. Multiple-choice questions accept between two and 255 options; the response includes the selected option and probabilities for every candidate. A scoring rubric produces a continuous value calculated from the probabilities assigned to its different levels.

These formats can be mixed within a single request. A call recording paired with account information could be used to identify the relevant department and assess urgency. Every question accesses the same content, but none can see the others’ answers. A decision that depends on an earlier result therefore requires additional orchestration in the program.

The context limit is 262,144 tokens for the entire request. That includes the shared content, question instructions, and criteria. The 1,000-question ceiling therefore does not represent 1,000 sequential reasoning steps or 1,000 analyses that each have access to that full capacity. Shared input costs, with latency measured under specific conditions

The announced price is $0.10 per million input tokens, with no output-token charges. Shared content is billed only once per request; each question adds its own instructions and criteria. The “nearly five times cheaper” comparison covers the seven shared evaluations against Gemini 3.5 Flash-Lite accessed through the Interactions API.

The reported end-to-end latency of 150–250 milliseconds applies to contexts of up to 16,000 tokens. In those tests, latency remains nearly flat between one and 100 questions. These measurements do not establish the same response time at the maximum of 1,000 questions or with a full context window. A routing component within larger products

Internally, Gemini Notebook uses these decisions to route requests to an appropriate model tier and reasoning level. The experimental CC service selects a reaction from more than 200 emojis. In Search, the API contributes to a layer that detects adversarial behavior involving MCP tools.

The Stitch Rapid experiment illustrates a creative application: product flows become interactive wireframes that can be edited by voice or touch. The demonstration shows the output of the complete product, which incorporates the decision API. It does not mean Gemini Decision 1 generates the screens or their code itself. Once a flow is selected, the screens can be transferred to the full Stitch environment for further design work.