Google has released the Gemini Decisions API, a programming interface designed for tasks where artificial intelligence needs to quickly evaluate data, select actions, and verify specified conditions. At the core of the new solution is the Gemini Decision 1 model, capable of analyzing multiple types of information simultaneously and processing up to 1,000 questions in a single request. According to the developers, this is not just another generative model, but a specialized tool for decision-making where the required output is an evaluation, choice, or numerical metric rather than a detailed response.
Decision in a Single Pass Instead of Sequential Text
The main difference between the new release and conventional generative models lies in how requests are processed. Instead of sequentially "writing out" text, Gemini Decision 1 evaluates the proposed options in a single pass. As a result, the median computation time is 50–100 ms, and taking full request processing into account, the latency increases to 150–250 ms. This approach makes the model suitable for scenarios where reaction speed is critical rather than the volume of generated text.
Multimodality and Three Evaluation Modes
The model works with text, images, audio recordings, video, and PDF documents, allowing different data types to be combined within a single context. For example, a phone call recording can be correlated with account information, and an insurance claim scan can be verified alongside vehicle damage photos. The maximum total context reaches 256,000 tokens. The developers have provided three evaluation modes: determining the probability of a statement's truth, selecting the most suitable option from a list, and scoring based on a predefined scale. Each result returns a confidence score from 0 to 1, enabling the application to automatically route doubtful cases to human review — in the demonstration interface, this looks like detecting a spilled liquid with a confidence of 0.95 and recommending the action "wipe with cloth" with a 98% probability.
Where Google Is Already Applying the New API
The company plans to use the Gemini Decisions API as an intermediate control layer for AI systems. For instance, Gemini Notebook uses it to select the appropriate model and reasoning level depending on request complexity: latency for routine requests is reduced by up to 40%, while deep analysis models are reserved for complex tasks. In the Stitch builder (formerly known as Galileo AI), the technology helps prototype interfaces, and Google Search uses it to detect malicious behavior when interacting with MCP tools.
Tests, Efficiency, and Pricing
Google provides comparative test results covering request routing, logical reasoning, ranking, and multimodal understanding. According to the developer, Gemini Decision 1 showed the best results in this test suite and minimal calibration error. Under identical test conditions, the model proved to be nearly five times more cost-effective than Gemini 3.5 Flash-Lite with API. Usage costs $0.10 per million input tokens, while output tokens are not billed: the general input data is paid for once regardless of the number of evaluated questions, and additional costs are formed solely by instructions and criteria for each check.
Developer Access
Developers can already obtain an API key in Google AI Studio and use the free tier to build prototypes. Integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, and other tools are rolling out gradually, and the company promises to add Gemini support in Google Cloud later. Thus, the Gemini Decisions API is positioned as a cheap and fast "decision layer" that can be embedded into existing AI pipelines without generating redundant tokens.
