Google has introduced the Gemini Decisions API — a new software interface designed to help AI agents instantly make predefined decisions without generating long text responses. The service runs on the basis of the new Gemini Decision 1 model.

What the new API can do

The main feature of the tool is that it is capable of evaluating up to 1000 independent questions for a single multimodal input (which may contain text, images, audio, video, and PDF files) in a single pass, reports Interesting Engineering.

Instead of generating text step by step token by token, the model evaluates ready-made options and returns probabilities and confidence indicators. Thanks to this, the latency is only 150–250 milliseconds from start to finish (while the median computation time of the model itself is 50–100 milliseconds).

Developers can use the API for limited tasks, in particular:

  • selection of tools and routing of requests to the necessary models;

  • verification of compliance with rules and policies;

  • context assessments;

  • simple yes/no checks, multiple-choice questions (up to 255 options), or evaluation based on predefined criteria.

    • The model supports a context window of up to 256,000 tokens and processes several types of data simultaneously without the need for additional third-party processors.

    Economy and competition in the market

    Google says Gemini Decision 1 is roughly five times cheaper than the Gemini 3.5 Flash-Lite model when using the standard Interactions API.

    The cost of using the Decisions API is $0.10 per million input tokens, with no charge for output tokens. Users pay for shared input data only once—regardless of how many questions are asked about it. A free tier is available in Google AI Studio for prototyping.

    The decision came amid growing competition: in October, OpenAI introduced its own similar product (the Decisions API based on the GPT-6 Luna model).

    Where the technology is already being used

    Google is actively integrating the new API into its own products to improve their performance.

  • Gemini Notebook: The API is used to route incoming requests to the appropriate model tiers, which reduces latency for everyday tasks by 40%.

  • Stitch platform: helps quickly process voice and sensor commands to create interactive wireframes.

  • CC Assistant: instantly selects emoji reactions from over 200 options.

  • Google Search: serves as a protective layer for detecting abuse when working with Model Context Protocol (MCP) tools.

    • The Decisions API is currently already available through Google AI Studio, and is also integrated into popular developer platforms, including OpenRouter, LangChain, Vercel, LiteLLM, and Pydantic AI.