THE ESSENTIALS IN 30 SECONDS
- One request can carry 1 to 1,000 independent yes/no, multiple-choice or ordered-rubric questions about text, images, audio, video or PDFs.
- Each answer comes with probabilities and a confidence value between 0 and 1, so applications can send uncertain cases to a fallback or a human.
- Google prices it at $0.10 per million input tokens with no charge for output; it says end-to-end latency is 150–250 ms.
Google has introduced the Gemini Decisions API, built around a new model called Gemini Decision 1 (gemini-decision-1). Announced on November 11, 2026 by Min Choi and Felix Yu of Google DeepMind, it is not a chat or text-generation endpoint: it answers structured questions about an input and returns probabilities instead of prose.
How it works
A developer sends one input, which may combine text, images, audio, video and PDFs, along with up to 1,000 independent questions evaluated in a single pass. The documentation lists three question types. Yes/no questions return the probability that a condition holds. Multiple-choice questions pick among 2 to 255 options and return a probability for each. Score questions grade the input on an ordered rubric of 2 to 255 levels, with 2 to 10 recommended, and return a weighted score.
Every answer also carries a confidence value from 0 to 1, defined in the docs as the estimated probability that the top answer is correct. Google's suggestion is that code can use it to route ambiguous cases to a slower model or a human reviewer. The context window is 262,144 tokens, shared between the input and all questions.
Price and speed
Only input is billed, at $0.10 per million tokens; output tokens cost nothing, and Google says there is no extra charge for thinking tokens, audio transcription or OCR. Shared input is counted once per request, plus each question's own instructions. A free tier is offered in Google AI Studio for prototyping. Google puts end-to-end latency at 150 to 250 milliseconds and says it stays nearly flat between 1 and 100 questions for inputs up to 16,000 tokens. It also claims the service is close to five times cheaper than Gemini 3.5 Flash-Lite on its own benchmark suite, where it reports the lowest calibration error.
Use cases and caveats
Google pitches the API for tool selection, model routing, policy checks and content grading. It says Gemini Notebook uses it to route queries, and that Google Search uses it as a safety layer to spot adversarial behaviour in MCP tools. Integrations with OpenRouter, LangChain, Vercel, LiteLLM and others are rolling out, while support in Google Cloud is described as coming soon.
The API sits on the v1beta version of the Gemini API, and neither the blog post nor the documentation says whether it is a preview or generally available. Accuracy and calibration claims come from Google's own tests.
