Google DeepMind has introduced the Gemini Decisions API, a new developer tool designed to help applications make fast, confidence-based decisions across text, images, audio, video, and PDFs. Powered by a new model called Gemini Decision 1, the API aims to make tasks such as AI model routing, content evaluation, policy checks, and tool selection faster and more cost-effective.

Announced on November 11, 2026, the Gemini Decisions API takes a different approach from traditional generative AI models. Instead of generating responses token by token, it evaluates predefined choices and returns probabilities and confidence scores in a single forward pass. This makes it particularly useful for applications that need to make quick, structured decisions rather than generate lengthy responses.

Developers can submit a shared multimodal input alongside as many as 1,000 independent questions in a single request. The API supports inputs of up to 256K tokens, allowing applications to evaluate multiple aspects of the same information without repeatedly sending the underlying content.

How the Gemini Decisions API works

At the heart of the new API is Gemini Decision 1, a purpose-built model for closed-set decisions. Rather than asking a general-purpose AI model to reason through a prompt and generate an answer, developers can define the possible outcomes and let the model score them directly.

Google says the API delivers 50–100 milliseconds of median model computation time, with end-to-end latency of approximately 150–250 milliseconds. Its performance also remains nearly flat when the number of questions increases from one to 100 in a request, based on Google’s reported benchmarks.

The API supports three types of decisions:

• Yes/No (noul): Estimates the probability that a particular condition is true or false. Developers can provide descriptions of what qualifies as either outcome.
• Multiple Choice (choice): Selects the most appropriate option from between two and 255 candidates, returning normalized probabilities for every available choice.
• Ordered Rubric (score): Evaluates an input against an ordered scoring rubric, returning a continuous weighted score alongside probabilities for individual levels.

Every answer includes a confidence score between zero and one. Applications can use these scores to determine whether to proceed automatically, send a request to another AI model, or involve a human reviewer when the result is uncertain.

For example, an application processing insurance claims could evaluate a scanned claim document alongside photographs of property damage. It could check whether the images show damage, select the appropriate claim category, and assign a severity score in one request.

Because the model natively handles multiple input formats, developers can combine text, images, audio, video, and PDF documents in the same evaluation. Image and video resolution can also be adjusted to balance visual detail against token consumption.

Designed for faster AI agents and model routing

One of the most interesting applications of the Gemini Decisions API is AI model routing. Not every request needs an expensive model with advanced reasoning capabilities. Simple questions can often be handled by smaller, faster models, while complicated tasks can be directed to more capable systems.

The new API provides a way to make those decisions quickly, helping applications select the appropriate model or tool without introducing the additional latency of a conventional generative AI call.

Google is already using the technology across several of its products.

Gemini Notebook uses the Decisions API to classify incoming requests and route them to an appropriate model tier and thinking level. Google says this approach can reduce latency for everyday queries by up to 40%, while reserving more advanced reasoning models for complex requests.

Google’s Stitch design tool is also using the API in a new rapid canvas experiment. It helps generate playable wireframes of product flows in seconds and supports almost instant edits through voice or touch interactions.

Meanwhile, CC from Google Labs uses the technology to select appropriate emoji reactions from a collection of more than 200 options. Google Search also uses the Decisions API as a safety layer for its MCP tools, helping detect adversarial behavior and potentially unsafe content.

These examples demonstrate how a relatively small, specialized decision-making model can support larger AI experiences without requiring a full generative response for every interaction.

Pricing, availability, and developer access

The Gemini Decisions API is available through the Gemini API in Google AI Studio, powered by the Gemini Decision 1 model, identified as gemini-decision-1.

Google has priced the model at $0.10 per 1 million input tokens, with no charges for output tokens. A no-cost API tier is also available for prototyping in Google AI Studio.

An important detail is that the shared input is billed only once per request, regardless of how many questions developers evaluate against it. Each additional question contributes only its own instruction and criteria tokens to the bill.

Google says the pricing structure eliminates output-token charges, variable thinking-token costs, and separate audio transcription or optical character recognition (OCR) costs.

Developers can get started by generating an API key through Google AI Studio. Google has also published implementation guides for Python, JavaScript, and REST, alongside demonstration applications including Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper.

The API is also gaining support across several third-party developer platforms, including OpenRouter, LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode. Support for Gemini in Google Cloud is coming soon.

With Gemini Decision 1, Google is targeting a part of AI development that doesn’t always get as much attention as bigger models and more impressive generative capabilities: making the right decision quickly. For developers building AI agents and multimodal applications, the ability to evaluate multiple conditions in one low-latency request could reduce both operating costs and unnecessary model calls.