Google launched the Gemini Decisions API on November 11, a service for applications that need to classify information and choose between predefined outcomes. The new Gemini Decision 1 model can evaluate text, photographs, audio recordings, video and PDF documents. Its responses are probabilities, choices and scores that software can use to route requests, assess documents or select a tool.
The product is accessible through the Gemini Developer API using the v1beta interface, and in Google AI Studio. Its model identifier is gemini-decision-1. A request can contain up to 1,000 independent questions within a total window of 262,144 tokens. That capacity includes both the shared information being evaluated and the questions and their criteria; the limit does not represent a number of words or documents.
Numerical results for bounded decisions
Three forms of evaluation are available. A yes-or-no check, called noul in the API, returns the probability of an affirmative answer, between 0.01 and 0.99. The choice type compares between two and 255 options. For an ordered rubric, score calculates the average of the level indices, weighted by their probabilities. Numbering starts at zero, and the score can fall between two levels.
Each evaluation includes a confidence indicator. For a choice, it corresponds to the probability of the option considered most appropriate; for a binary question, it reflects the more likely of yes and no. Google recommends setting thresholds using labeled examples from the application's traffic. In its support examples, the result can send a request to a work queue or to a person for review, depending on the program's criteria.
How the evaluation result is used
- 1 · Shared information
Text, documents and media evaluated against application criteria - 2 · Numerical evaluation
Probability, choice or score, together with model confidence - 3 · Application decision
Automatic routing or human review according to configured thresholds
Gemini Decision 1 evaluates the options in a single model pass and produces zero text output tokens. Its specification separates the product from models that write explanations, generate code or make function calls. Selecting a tool produces a result delivered to the application; executing that tool belongs to the program receiving the response. Web search, code execution and image generation are not capabilities of this model.
Documents and media in the same request
The shared information can combine several content types. For example, Google's documented demonstrations pair order data with a customer's message or check elements within a PDF. For files sent inline, the request ceiling is 100 MB, with a 50 MB limit for PDFs. Accepted HTTPS media links have a limit of 15 MB for each external resource.
Larger or reusable files can be uploaded through the Files API and then referenced in a request. The file service allows a maximum of 2 GB per file and 20 GB per project. Uploaded files are retained for 48 hours and can be deleted earlier. This temporary storage has its own limits, separate from the number of tokens the model can analyze in an evaluation.
Text and visual material contribute to token consumption, the units through which the API measures the information it processes. For images and video, the selected resolution changes the detail analyzed and the corresponding consumption. The Decisions API documentation specifies that ultra-high resolution applies only to images. Audio blocks and documents do not use this setting in requests to the new service, even though other Gemini models have different media controls.
Input costs and project-level limits
The published standard price is $0.10 per million input tokens, whether the information comes from text, images, audio, video or PDFs. Shared information is counted once within a request; each question adds its own instructions, criteria and formatting elements. There are no billable output tokens. Context caching and processing through the Batch API are not available for Gemini Decision 1.
Google also lists a free tier for prototypes. For the paid Tier 1, standard limits are 1,500 requests per minute, one million tokens per minute and 50,000 requests per day. These limits apply to the project, rather than to each API key separately. Usage is checked against each dimension, so reaching the token limit can stop requests before the numerical request ceiling is reached.
In measurements published by the company, median end-to-end latency is 150–250 milliseconds, while model computation takes 50–100 milliseconds. The results presented cover contexts of up to 16,000 tokens and between one and 100 questions in a request. This testing range is distinct from the API's maximum supported capacity.
Business access and application integration
Google provides official Google GenAI libraries, including for Python and JavaScript/TypeScript. Examples for the new service require SDK 3.2 or later and also document REST access. Authentication uses a key associated with a Google Cloud project. Authorization keys are bound to a service account, while the project manages billing, collaborators and permissions.
Romania is listed among the supported regions for Google AI Studio and the Gemini API. The service terms require paid services for applications made available to users in the European Economic Area, Switzerland or the United Kingdom. In those regions, the paid-service data-use rules also apply to AI Studio and free quota: Google states that inputs and responses are not used to improve its products.
The launch concerns a specialized model for integration into applications, with questions and outcomes bounded by the developer. Initial access is offered through the Gemini Developer API and AI Studio; the announcement describes support in Gemini on Google Cloud as a future stage. Integrations with other ecosystem tools are rolling out.
