Google's Gemini Decision 1 answers up to 1,000 questions in one pass

Google DeepMind released Gemini Decision 1 on 11 November, a model that picks an answer from the options you supply. It powers the new Gemini Decisions API, which takes up to 1,000 closed questions about one input and returns all the answers in 150 to 250 milliseconds.

How it differs from a chat model

A regular LLM writes its answer token by token. Gemini Decision 1 scores the candidate options you give it in one forward pass and returns a probability for each. The API supports three question types:

• yes or no
• multiple choice with 2 to 255 options
• a score on an ordered rubric

The input can be text, images, audio, video or PDF files, up to 256,000 tokens. In Google's own example, the model looked at a photo of a kitchen counter and in one pass answered whether there was a stain, which cleaning action to take and how bad the mess was.

“Every answer includes a confidence score (from 0 to 1) so your code can route ambiguous cases to a fallback.”

— Min Choi and Felix Yu, Google DeepMind, Google blog announcement of the Gemini Decisions API

Quote source: Google, company blog, 11 November 2026.

With that score you can split the work: let the cheap model handle the clear cases and send the uncertain ones to a larger model or to a person. Google says the model has the lowest calibration error in its test suite, so its stated confidence matches how often it is right, but the blog post gives no benchmark figures.

Price and where Google already uses it

Input costs $0.10 per million tokens, and Google charges nothing for output. If you send 1,000 questions about one document, you pay for the document once and for each question's own instructions. Google puts the cost at about a fifth of Gemini 3.5 Flash-Lite on the same tests.

Inside Google the model runs in four products:

• Gemini Notebook, where it routes queries between models and cut routing latency by up to 40%
• Stitch, for quick wireframe edits
• Google Search, as a safety layer that checks MCP tools for adversarial behaviour
• CC from Google Labs, to pick an emoji reaction from more than 200 options

The API is open in Google AI Studio and the Gemini API, with a free tier for prototypes. Integrations with OpenRouter, LangChain, Vercel and other tools are rolling out, and Google Cloud support is "coming soon".

A crowded category

Google arrived after others. The Decoder notes that the startup Jev introduced a similar model first and that OpenAI, Cloudflare and Microsoft followed with their own. Routing requests between models is a business of its own by now: in August Stripe agreed to buy OpenRouter.

If you build an AI agent that sorts or routes requests, test Decision 1 on your own labelled data before you swap out a chat model. It answers only from the options you list, so open questions still need a generative model, and Google has published no benchmark numbers you could check its claims against.

Google described the model, its prices and its internal use, and we have not tested it. This report covers a product launch and is not a recommendation to buy any service.