Google has introduced the Gemini Decisions API, a new Gemini API endpoint intended for applications that need fast, bounded decisions rather than open-ended generated responses. The service is powered by Gemini Decision 1, a model built to score candidate options across text, images, audio, video and PDFs.
The launch reflects a growing need for AI systems that can make operational choices inside larger software workflows: selecting tools, routing requests, checking policy rules, ranking alternatives or grading context. In those cases, a developer may need a probability and confidence score more than a paragraph of explanation.
Designed for closed-set choices
Google describes the API as a tool for yes-or-no questions, multiple-choice decisions and rubric-based scoring. A developer can submit one multimodal input and ask up to 1,000 independent questions about it in a single request.
The system supports inputs spanning text, images, audio, video and PDFs, with contexts of up to 256,000 tokens. That combination is meant to let one service evaluate a broad range of material without requiring developers to build separate decision pipelines for each media type.
Examples given by Google include tool selection, policy checks, model routing and context grading. A system could use the endpoint as a gatekeeper before calling another model or service, or as a scoring layer to determine which action should be taken next.
Probabilities instead of generated answers
The core technical distinction is that Gemini Decision 1 scores candidate options in a single forward pass. It does not generate an answer token by token in the way a conversational model normally does.
According to Google, the endpoint returns probabilities and confidence scores. This creates a more structured output for applications that need to make a defined choice and also need a measure of how certain the model is about that choice.
For production AI systems, that confidence signal can be important. It can help developers set thresholds, send uncertain cases for review or choose a fallback route when the model’s score does not clear a required level. The API does not eliminate the need for evaluation, but it offers a decision-oriented primitive that is more directly aligned with those tasks.
Latency and pricing claims
Google says Gemini Decision 1 has model-compute latency of 50 to 100 milliseconds and an end-to-end latency of 150 to 250 milliseconds. It says this performance remains nearly flat from one to 100 questions per request when working with multimodal inputs and contexts up to 16,000 tokens.
The company lists pricing at $0.10 per million input tokens, with no billed output tokens. Because the model scores options rather than emitting a conventional text response, the pricing structure is tailored to a workload where output-token generation is not the central task.
Google also says the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API across its shared benchmark suite. As with all vendor benchmark claims, the practical result for a given deployment will depend on the inputs, questions, evaluation rules and surrounding system design.
Calibration becomes a product feature
Google positions calibration as a central feature of the new model. A calibrated system aims to make confidence values meaningful: high-confidence answers should be more likely to be correct than low-confidence answers. That property is especially relevant when an AI output is used to trigger a workflow, enforce a rule or route a user request.
The company says Gemini Decision 1 leads its benchmark suite across agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. The announcement does not make those results a substitute for developers testing their own data and edge cases, but it signals Google’s emphasis on decision quality as well as raw answer accuracy.
For developers, the practical question is not simply whether a model can choose an option. It is whether its scores can be used reliably enough to support automation. A model that can rank alternatives but cannot indicate uncertainty clearly is harder to place safely inside a consequential workflow.
A new building block for AI agents
The API arrives as companies assemble AI agents from multiple specialized components. In such systems, a language model may interpret a request, a decision model may choose the next tool, and another service may perform the action. Fast scoring can therefore be useful even when the user never sees a model response from the decision layer.
Google’s offering is available through the Gemini API in Google AI Studio, where developers can access an API key and documentation for Python, JavaScript and REST. The launch gives developers a purpose-built option for closed-set multimodal decisions, especially where response speed, cost and confidence reporting matter as much as fluent text generation.
Its importance will depend on whether the promised calibration and latency hold up in real-world workloads. But the product framing is clear: rather than treating every AI task as a conversation, Google is offering a model endpoint intended to make many small, explicit decisions quickly.
This article is based on reporting by Google AI Blog. Read the original article.
Originally published on blog.google
