Gemini Decisions API: Fast, calibrated multimodal decisions
On November 11, 2026, Google announced the launch of the Gemini Decisions API, powered by the new Gemini Decision 1 model. This API allows users to evaluate up to 1,000 yes/no, choice, and rubric-scoring questions rapidly across text, images, audio, video, and PDFs.
The Gemini Decisions API is designed for fast, closed-set decisions and is accessible through the Gemini API in Google AI Studio. It offers specialized capabilities for bounded choices, including tool selection, policy checks, model routing, and context grading. Users can submit a single multimodal input of up to 256K tokens along with multiple independent questions.
Unlike traditional models that generate answers token by token, Gemini Decision 1 scores candidate options in a single forward pass, providing probabilities and confidence scores with an end-to-end latency of 150-250 milliseconds at a cost of $0.10 per million input tokens, with no charges for output tokens.
Performance and Calibration
Gemini Decision 1 demonstrates strong performance across various benchmarks, including agent routing, reasoning, ranking, and multimodal understanding, achieving the lowest calibration error. The model maintains median latency of 50-100 milliseconds for compute, resulting in overall latency of 150-250 milliseconds across multimodal inputs and contexts of up to 16K tokens.
With zero output tokens billed, the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite when evaluated against shared benchmarks.
Multimodal Evaluation Capabilities
Built on Gemini’s native multimodal architecture, Gemini Decision 1 can process various input types directly in a single pass. Users can combine modalities, such as pairing audio recordings with text notes or images with PDFs, and adjust resolution for image and video inputs to manage visual detail and token costs.
The API supports multiple question types in a single request, including yes/no checks, multiple choice, and rubric scores, with each answer accompanied by a confidence score. The question types include:
- noul (Yes/No): Estimates the probability of a yes/no condition with optional description criteria.
- choice (Multiple Choice): Selects from 2 to 255 choices, returning the most likely option and probabilities for each candidate.
- score (Ordered Rubric): Grades inputs based on an ordered rubric, providing a continuous weighted score and probabilities for each level.
For example, users can check a photo of a breakroom counter alongside a text note against all three question types in one request using the API.
Applications of the Decisions API
Teams at Google are already utilizing the Decisions API for various applications, including:
- Dynamic model routing: Gemini Notebook employs the Decisions API to route incoming queries to the optimal model tier, reducing latency by up to 40% for everyday queries.
- Rapid design iteration: The rapid canvas experiment in Stitch uses the API to create playable wireframes quickly and apply edits in real-time.
- Emoji selection: CC from Google Labs utilizes the API to select appropriate emoji reactions from a large palette instantly.
- Safety layer in Search: Google Search leverages the Decisions API to detect adversarial behavior and content effectively.
Getting Started with the Decisions API
Gemini Decision 1 is priced at $0.10 per million input tokens, with no charges for output tokens. An API no-cost tier is available for prototyping in Google AI Studio. Users are billed once per request for the shared input, regardless of the number of questions evaluated.
Developers can begin using the API by obtaining an API key in Google AI Studio and exploring demo applications such as Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper. Comprehensive guides for Python, JavaScript, and REST are available in the Decisions API documentation.
Integrations with various ecosystems are underway, including OpenRouter, LangChain, and Vercel, with additional support coming soon to Google Cloud.
Source: Google AI Blog