Google Unveils Gemini Decisions API for Rapid Decision-Making

On November 11, Google introduced the Gemini Decisions API based on the Gemini Decision 1 model. Unlike typical generative systems, the service does not generate responses but evaluates predefined action options and returns their probabilities. Developers can use it for distributing requests among AI models, checking rules, and selecting tools in agent operations.

How Gemini Decision 1 Works

In a single call, the API accepts a general dataset and up to 1,000 independent questions. The model analyzes text, images, audio, video, and PDFs within a context of 262,144 tokens. It supports three types of decisions: yes/no answers, multiple-choice selection, and evaluation on a given scale. Each result is accompanied by a confidence score, allowing contentious cases to be directed for further review.

Speed and Cost of the New API

Google has set a price of $0.10 per million input tokens, without charging for output tokens, as the model does not generate them during operation. General input data is counted once, but instructions for each question add to token usage. According to the company's own measurements, with a context of up to 16,000 tokens, processing took 150–250 ms from request to response, and the latency remained almost unchanged when the number of questions increased from 1 to 100. These figures do not guarantee the same speed for any data volume.

Test Results and Service Availability

In the JevBench trial, the model scored 92.2%, with a probability calibration error of 0.0250, according to Google data. However, in the When2Call test, it lagged behind TypeSafe Jev 1.14, so it cannot be said to excel in all tasks. Within Google, the new API is already used for routing requests in Gemini Notebook, where the company reports a reduction in everyday request latency by 40%. Developers have access to a preliminary version via the Gemini API and Google AI Studio, including a free plan for testing.