Gemini Decisions API

The Gemini Decisions API is a new tool designed to facilitate quick, probabilistic decision-making across various data types, including text, images, audio, video, and PDFs. This innovative solution calculates confidence scores for numerous questions simultaneously, offering an economical approach to query routing, content grading, and policy rule verification.

Developers can begin integrating this API by obtaining an API key in Google AI Studio and consulting the available documentation for Python, JavaScript, and REST protocols.

The Gemini Decisions API introduces a dedicated endpoint for expedited, closed-set decisions, leveraging the Gemini Decision 1 model (gemini-decision-1), which is accessible through the Gemini API within Google AI Studio. This API provides specific functionalities tailored for bounded choices such as tool selection, policy verification, model routing, and context grading.

Users can input a single multimodal query, encompassing text, images, audio, video, and PDFs, with a token limit of 256K. It can accommodate up to 1,000 independent questions in one request.

Unlike traditional generative models that generate answers token by token, Gemini Decision 1 analyzes candidate options in a single forward pass, delivering probabilities and confidence scores within approximately 150-250 milliseconds. This efficiency costs $0.10 per 1 million input tokens, with no additional charges for output tokens. The API's benchmark performance on agent routing, reasoning, ranking, and multimodal understanding is superior, boasting the lowest calibration error among its peers.

Gemini Decision 1's design emphasizes speed and cost-effectiveness. It processes options without decoding each token, maintaining median latency at 50-100 milliseconds of actual model computation. This efficiency spans multimodal inputs and contexts up to 16K tokens, remaining consistent even when handling up to 100 questions per request. Additionally, the API does not bill for output tokens, making it approximately five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API across comparable benchmarks.

One of the key features of the Gemini Decisions API is its ability to process multimodal inputs directly within a single forward pass. This means users can combine various data types, such as audio recordings with account metadata or scanned PDF claims with corresponding images, and adjust image and video resolutions to balance visual detail and token usage.

Each answer generated by the API includes a confidence score ranging from 0 to 1, allowing developers to route uncertain cases to fallback mechanisms or human reviewers.

The input format for the Gemini Decisions API mirrors that of the Interactions API. An illustrative example demonstrates how to evaluate a photo of a breakroom counter alongside a text note against three different question types in a single request. Multiple major tech companies, including Google, are already utilizing the Decisions API as a low-latency control plane, streamlining product development by defining options instead of training custom classifiers or waiting for generative responses.

Google has also introduced a no-cost tier for API access, allowing developers to prototype applications in Google AI Studio without upfront costs. The API operates on a pay-as-you-go model, billing only for input tokens and not for output tokens, ensuring cost efficiency. Developers can begin building applications by creating an API key in Google AI Studio, exploring demo apps for real-world scenarios, and following the provided guides in the Decisions API documentation.

Ecosystem integrations are expanding to include platforms like OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, with further support for Gemini in Google Cloud.