Google launches Gemini Decisions API for fast multimodal choices

A new tool for bounded choices

Google has released the Gemini Decisions API, a specialized endpoint designed for making fast, closed-set decisions. Unlike standard generative models that create text word by word, this API is built for specific tasks like choosing a tool, checking a policy, or routing a request to a different model. It uses a new model called gemini-decision-1 to provide these results efficiently.

The API is designed to handle multimodal inputs, which means it can process different types of data simultaneously. A single request can include text, images, audio, video, or PDFs up to 256K tokens in size. This allows developers to combine different types of information, such as pairing a scanned PDF claim with photos of damage, to reach a conclusion.

One of the primary goals of this release is to provide purpose-built primitives for bounded choices. These are situations where the AI must pick from a set of pre-defined options rather than creating an open-ended response. By focusing on these specific types of queries, the system achieves much higher speed and better calibration than general-purpose models.

How the scoring system works

The Gemini Decisions API operates by scoring candidate options in a single forward pass. This means the model looks at the input and all questions at once to provide probabilities and confidence scores. Because it does not use token-by-token decoding, the median model compute time is kept between 50-100 ms, resulting in a total end-to-end latency of 150-250 ms.

Developers can ask up to 1,000 independent questions in a single request. The API supports three specific types of evaluations to handle different needs. These include 'noul' for yes/no questions, 'choice' for multiple-choice selections between 2 and 255 options, and 'score' for grading inputs against an ordered rubric.

Every answer provided by the model includes a confidence score ranging from 0 to 1. This allows developers to build systems that know when they are unsure. If a confidence score is too low, the code can automatically route that specific case to a human reviewer or a different fallback process to ensure accuracy.

Cost and performance benefits

The pricing structure for the Gemini Decisions API is designed to be highly economical for high-volume tasks. It costs $0.10 per 1M input tokens, and notably, there are zero costs for output tokens. This makes it nearly 5x cheaper than using the Gemini 3.5 Flash-Lite model with the Interactions API for the same benchmark suite.

The API also simplifies billing for complex multimodal requests. The shared input is billed only once per request, regardless of how many questions are being evaluated. Each question only adds the cost of its own instructions and criteria tokens. This structure removes extra costs for things like audio transcription or OCR, which are often separate charges in other systems.

In terms of performance, Gemini Decision 1 leads across several benchmarks including agent routing and reasoning. It maintains a nearly flat latency even when the number of questions per request increases from 1 to 100. This consistency makes it a reliable choice for applications that require rapid, high-frequency decision-making without unpredictable spikes in processing time.

Real-world applications at Google

Several teams within Google are already using the Decisions API as a low-latency control plane. For example, Gemini Notebook uses it as a real-time router to direct queries to the best model tier. This helps serve everyday queries with up to 40% lower latency while saving complex reasoning models for more difficult tasks.

Other internal projects demonstrate the API's versatility in design and safety. The Stitch project uses it to generate playable wireframes of product flows in seconds and to apply voice or tap edits almost instantly. Meanwhile, Google Search uses the API to detect adversarial behavior and content against specific evaluations within a very tight latency window.

Smaller interactions also benefit from this speed. A project called CC from Google Labs uses the API to pick the correct emoji reaction from a palette of over 200 options. Because the API returns the winner and normalized probabilities in a fraction of a second, the reactions appear instantly to the user without adding any noticeable delay to the interface.

Availability and integration

The Gemini Decisions API is now available through the Gemini API in Google AI Studio. Developers can start prototyping immediately using a no-cost tier. Google has provided several demo apps, including Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper, to help users see the model in action.

The company is also rolling out ecosystem integrations to make the tool accessible to a wide range of developers. Support is being added to popular frameworks like LangChain, Vercel, LiteLLM, and Pydantic AI. These integrations allow teams to plug the Decisions API into their existing workflows more easily than building from scratch.

For those ready to build, Google provides documentation and guides for Python, JavaScript, and REST. The API also supports the same Content block format as the Interactions API, making it easier for teams to transition existing multimodal projects. Support for Gemini in Google Cloud is also expected to arrive soon for enterprise users.