Google has launched the Gemini Decisions API and Gemini Decision 1 model for rapid, closed-set judgments across multimodal inputs. It can answer up to 1,000 questions per request with confidence scores.
Google has launched the Gemini Decisions API, a new endpoint designed to make rapid, closed-set judgments—such as yes/no checks, multiple-choice selections, and rubric-based scoring—across a wide range of media. The service is powered by a new model, Gemini Decision 1 (gemini-decision-1), and is available through the Gemini API in Google AI Studio.
The central promise is speed and scale: developers can submit a single multimodal “shared input” (text, images, audio, video, or PDFs) and ask up to 1,000 independent questions about that same material in one request. Google positions the API as a practical tool for tasks like policy compliance checks, routing requests to the right tool or model, grading or ranking content, and “context grading” for downstream systems.
Unlike generative models that produce text sequentially, token by token, Gemini Decision 1 evaluates the provided answer options in a single forward pass. The result is a set of probabilities and confidence scores for each question rather than long-form explanations. Google says the API returns results with 150–250 milliseconds end-to-end latency, with model compute typically in the 50–100 millisecond range. Performance is intended to stay nearly consistent as developers increase question count from 1 to 100 in a request, and the system supports multimodal contexts up to 16,000 tokens; the broader input limit for the endpoint is up to 256,000 tokens.
Google also emphasizes calibration—how closely confidence scores match real-world correctness. On internal benchmarks spanning agent routing, reasoning, ranking, and multimodal understanding, the company says Gemini Decision 1 leads the suite and posts the lowest calibration error, a metric important for building systems that can automatically defer uncertain cases to humans or fallback logic.
Pricing is structured to favor high-volume decision workloads. Gemini Decision 1 costs 0 charged for output tokens. Because the endpoint returns scores rather than generated text, there are no output-token fees, and Google says there are no variable “thinking token” charges either. The shared input is billed once per request regardless of how many questions are asked, while each question adds only its own instruction and criteria tokens. Google also notes that the API does not impose separate charges for audio transcription or OCR when working with audio or scanned documents.
The multimodal design allows developers to mix formats in one request—for example, combining a call recording with account metadata, or pairing a scanned PDF claim with photos of damage. For image and video, resolution can be adjusted to control visual detail and token costs. Each response includes a confidence score between 0 and 1, intended to make it easier to build workflows that automatically route borderline decisions to a secondary system or to human review.
Google says teams inside the company are already using the Decisions API as a low-latency “control plane,” enabling new product features by defining decision options rather than training custom classifiers or waiting on full generative responses. One cited example is Stitch Rapid, which Google says uses the API to generate an entire product flow in seconds.
Developers can start using the tool through Google AI Studio by generating an API key and following documentation for Python, JavaScript, and REST. Google is also rolling out ecosystem support via integrations including OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode, and the google-genai SDKs. Demo apps highlighted by Google include Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper, and a no-cost tier is available for prototyping in AI Studio.
Why it matters
Many AI-powered products spend time and money generating text when they really need fast, reliable decisions—such as “allow/deny,” “which tool,” or “what score.” A calibrated, multimodal scoring endpoint could lower latency and cost for content moderation, customer support triage, and agent routing, while enabling safer automation through confidence-based escalation. The next test will be how well Gemini Decision 1’s benchmark claims translate to real-world workloads and how quickly third-party tooling standardizes around the new endpoint.