Gemini Decisions API: Fast, calibrated multimodal decisions

Google has unveiled the Gemini Decisions API, a new tool designed to enable swift, probabilistic choices across various media types, including text, images, audio, video, and PDF documents. This API computes confidence scores for multiple questions in a single operation, offering a cost-effective solution for tasks such as routing queries, grading content, or verifying policy rules.

Developers can begin utilizing the API by obtaining an API key through Google AI Studio and consulting the provided documentation for Python, JavaScript, and REST implementation guidelines.

The Gemini Decisions API introduces the Gemini Decision 1 model (gemini-decision-1), which is specifically engineered for closed-set decisions. This model supports multimodal inputs, allowing up to 256K tokens encompassing text, images, audio, video, and PDFs in a single request. The API can accommodate up to 1,000 independent questions related to a single multimodal input.

Instead of generating answer text token by token, the Gemini Decision 1 model calculates scores for candidate options in one forward pass, delivering probabilities and confidence scores within a latency range of 150-250 milliseconds. This approach significantly reduces costs and latency compared to traditional methods that generate text outputs sequentially.

Benchmark tests reveal that Gemini Decision 1 outperforms its counterparts in various aspects, such as agent routing, reasoning, ranking, and multimodal understanding, exhibiting the lowest calibration error among similar models. Notably, Gemini Decision 1 scores options without the need for token-by-token decoding, maintaining a median latency of 50-100 milliseconds of model compute (150-250 milliseconds end-to-end), even when handling up to 16K tokens in context and multiple questions per request.

The absence of billing for output tokens makes Gemini Decision 1 approximately 5 times more cost-effective than Gemini 3.5 Flash-Lite with the Interactions API, according to shared benchmark comparisons.

One of the key features of Gemini Decision 1 is its ability to process text, images, audio, video, and PDF documents directly in a single forward pass, leveraging its native multimodal architecture. This capability enables users to combine multiple modalities within a single input, such as merging a call recording with account metadata or attaching damage photos to a scanned PDF claim. Users can also tailor the image and video input resolutions to balance visual detail against token costs.

Further flexibility is offered through the ability to pose a mix of question types—such as yes/no checks, multiple-choice inquiries, and rubric scoring—against the same shared input within a single API call. Each answer returned by the API includes a confidence score ranging from 0 to 1, empowering developers to direct ambiguous responses to fallback mechanisms or human review. The input field adheres to the same Content block format as the Interactions API, ensuring compatibility and ease of use.

Google teams across various departments are already leveraging the Decisions API as a low-latency control plane to expedite feature development. For instance, Stitch Rapid utilizes the Decisions API to design product flows in seconds, achieving a fivefold increase in speed compared to traditional methods. Pricing for the Gemini Decision 1 API is set at $0.10 per 1 million input tokens, with no charges incurred for output tokens.

The API includes a no-cost tier accessible via Google AI Studio, facilitating prototyping and experimentation. The shared input is billed once per request, regardless of the number of questions evaluated, without additional costs for tokenized thinking or separate costs for audio transcription or OCR services.

To get started, developers can sign up for an API key in Google AI Studio, experiment with sample applications like Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper, and follow the available guides in the Decisions API documentation for Python, JavaScript, and REST integration. Additional ecosystem integrations are being rolled out for popular platforms such as OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, with support for the Google Cloud Gemini service to follow.