On November 11, 2026, Google released the Gemini Decisions API, built around a new model called Gemini Decision 1 (gemini-decision-1). It doesn't write words: hand it up to 1,000 closed-ended questions at once and it judges them all in a single forward pass, returning only each question's probability and confidence score, end to end in 150 to 250 milliseconds.
What It Can Judge in One Go
Three question types: yes/no, multiple choice, and rubric scoring. A single request carries one multimodal input—text, images, audio, video, PDFs, up to about 262,000 tokens—plus up to 1,000 independent questions. The input is billed once, so the more questions you ask, the better the deal.
Google positions it narrowly: tool selection, policy checks, model routing, and content grading. Gemini Notebook already uses it to decide in real time which model tier and thinking level to route to, and Search uses it internally as a safety layer for MCP tools. Pricing is $0.10 per million input tokens, with output tokens free—because it produces no output tokens at all. Access runs through the Gemini API in Google AI Studio, with a free tier for prototyping.
How to Use It in Practice
Decide what you're judging before writing the questions: for a customer-service conversation, ask "does this involve a refund?", "is the sentiment angry?", "should this go to a human?"—three questions answered in one call, far cheaper than querying a chat model three times. Don't blindly trust low-confidence answers: set a threshold and send anything below it to a human or a larger model for a second look. Bulk grading, bulk labeling, and bulk routing are where it pays off most; asking just one or two questions at a time wastes its strength.
Know the Boundaries First
First, it doesn't generate free-form text—don't use it to write articles, summarize, or answer open questions. Second, the speed and benchmark numbers are Google's own: 150 to 250 milliseconds applies to inputs up to 16K tokens and 1 to 100 questions per request; Google gives no figure for larger inputs or 1,000 questions, and as of release there is no independent replication. Third, it currently runs only through the Gemini Developer API's v1beta—context caching and the Batch API aren't supported yet. OpenAI has a similar product; for comparison see OpenAI Decisions API is generally available: small decisions 10x faster and billed on input tokens only, and for Google's full announcement see Gemini Decisions API is here: judge 1,000 questions at once, results in 150 milliseconds.
