Google Launches Gemini Decisions API for Rapid, Multimodal Decision-Making

On November 11, 2026, Google unveiled the Gemini Decisions API, a new tool designed to facilitate rapid probabilistic decision-making across multiple data types, including text, images, audio, video, and PDFs. This API leverages the Gemini Decision 1 model to evaluate up to 1,000 independent yes/no, multiple choice, or rubric-scoring queries simultaneously through a single forward-pass computation, significantly reducing processing time and cost.

The core innovation of the Gemini Decisions API lies in its ability to provide confidence scores for each query without performing token-by-token generation typical in conventional large language models. This scoring mechanism achieves end-to-end latency between 150 and 250 milliseconds and costs $0.10 per one million input tokens, with no charges for output tokens, making it an efficient solution for decision routing, content grading, and policy compliance checks.

Gemini Decision 1 demonstrates strong performance and calibration across benchmarks involving agent routing, reasoning, ranking, and multimodal understanding, often outperforming previous models. It supports inputs combining multiple modalities in one request—for example, associating a call recording with metadata or pairing a scanned PDF with images—processing up to 256,000 tokens with minimal latency increases.

The API supports three decision types: yes/no classifications with probability estimates; multiple-choice selections between two and 255 options with normalized probabilities; and ordered rubric scoring that provides weighted scores with per-level confidence. Developers can tune input resolutions and utilize confidence scores to route ambiguous cases for further review or human intervention.

Various Google teams have already integrated the Decisions API into products such as Gemini Notebook, where it dynamically routes queries to optimal model tiers, reducing latency by up to 40%; Stitch Rapid, which uses it for instant wireframe generation and editing; CC from Google Labs, where it selects emoji reactions in real time; and Google Search, which employs it as a safety layer to detect adversarial content efficiently.

Pricing is competitive, with users charged based solely on input tokens, and a no-cost tier available for prototyping. Ecosystem integrations are expanding rapidly, with support forthcoming on platforms like OpenRouter, LangChain, and Gemini in Google Cloud. Comprehensive documentation and sample applications enable developers to start building immediately.

While the Gemini Decisions API excels in speed and cost efficiency, limitations include its bounding to predetermined question formats and dependence on confidence scores to handle ambiguous cases, necessitating fallback mechanisms. The model currently processes up to 256K tokens per request, which may constrain some large-scale or highly detailed workflows.