Gemini Decision 1 Release

Google released Gemini Decision 1 (gemini-decision-1) and a new Decisions API on Nov 11, 2026, for developers in the Gemini API and Google AI Studio. It is not a chatbot: it writes no text at all. Code sends one input (text, images, audio, video or PDFs) with up to 1,000 questions, and gets back a probability and a confidence score for each answer, which Google pitches for routing requests, checking content against rules and grading against a rubric. It costs $0.10 per million input tokens, with no output tokens billed and a free tier.

What it does

From Google's announcement and the Decisions API guide:

• Three kinds of question. Yes or no (noul, a probability from 0.01 to 0.99), a choice among 2 to 255 options (the winner plus every option's probability), and a score against 2 to 255 ordered rubric levels (a weighted score plus each level's probability). Every answer carries a confidence from 0 to 1, so code can send unsure cases to a person or a bigger model.
• Many questions, one pass. All the questions run against the shared input in a single forward pass and cannot see each other. Google says latency stays nearly flat from 1 to 100 questions: 50 to 100 ms of model compute, 150 to 250 ms end to end, for inputs up to 16,000 tokens.
• Mixed media in one input. A call recording with account notes, or a scanned claim with damage photos, in the same content format as Google's Interactions API.
• Google's own use. Google says Gemini Notebook already uses it to pick the model tier for each query, with up to 40% lower latency on everyday queries; Stitch's new rapid canvas, the CC app from Google Labs (choosing among more than 200 emoji) and Google Search (screening tool calls for adversarial content) use it too.

Google claims Gemini Decision 1 leads its own benchmark suite of agent routing, reasoning, ranking and multimodal tasks with the lowest calibration error, and that it is nearly five times cheaper on that suite than Gemini 3.5 Flash-Lite. Google has not published the suite's scores as a table that can be checked.

What it costs

Source: Gemini API pricing, read Nov 11, 2026. The shared input is billed once per request however many questions are asked; each question adds only the tokens of its own instructions and criteria. At that rate, 10,000 checks of a 2,000-token message cost about $2.

Limits

The model page lists 262,144 input tokens and up to 1,000 questions per request, and the API is called on the v1beta endpoint with Google's SDK 3.2 or later. The Batch API is not supported. Rate limits, read Nov 11, 2026:

Google's announcement names no country limits beyond the Gemini API's own, and says nothing about Google Cloud. Outside Google, it says integrations are rolling out on OpenRouter, LangChain, Vercel's AI SDK, LiteLLM and Pydantic AI. It is a developer model only: nothing in the Gemini app changes.

The demo

Google posted no launch video on YouTube with the announcement; the post shows a screen recording of Stitch's rapid canvas instead. Google's three sample apps in AI Studio, Doorstep Parcels (sorting delivery photos), Speed Sketch and Find the Zoo Keeper, run the model in the browser for anyone signed in to Google AI Studio.

What creators say

3:09 Rithesh Sreenivasan, who reviews new AI models for 51.8K subscribers, reran his 56 text and image questions with Gemini Decision 1 beside two rival decision models, D1 Ultra and Jev. Both rivals got 54 right and Gemini got 52, so in his words "on text and images, Google didn't win", though it was the only one to solve his robot grid route. He found its confidence numbers less reliable than D1 Ultra's (a calibration error of about 0.07 against 0.035) and advises tuning any cutoffs for this model.

7:39 He then sent 1, 10, 100, 500 and 1,000 questions about one price list in a single call: every answer was right, 1,000 questions took about 3 seconds and the call cost about half a cent. His verdict: D1 Ultra for text and images only; Gemini Decision 1 for audio, PDFs, CSV files, very long inputs or more than 128 questions a call, where it got 39 of his 42 new questions right, all three misses on video.

What we think

We think Gemini Decision 1 reaches Google Cloud's enterprise platform before Mar 1, 2027. Google brought its other Gemini API models there, the Decisions API is aimed at the routing and policy checks that large companies run at volume, and the launch post already lists five outside integrations but no Cloud date.

Confidence: Likely. Check by: Mar 2027.

We think OpenAI or Anthropic offers its own API that returns calibrated probabilities instead of text by Jun 30, 2027. Google's case for it is cost and latency (no output tokens, one pass for 1,000 questions), the two things developers who route traffic between models pay for most, and the post shows Google using it to route its own Gemini Notebook queries.

Confidence: Possible. Check by: Jun 2027.

Sources

• Gemini Decisions API: Fast, calibrated multimodal decisions, Min Choi and Felix Yu, Google, Nov 11, 2026
• Gemini API changelog, Google, Nov 11, 2026 entry
• Gemini Decision model page and Decisions API guide, Google, read Nov 11, 2026
• Gemini API pricing and rate limits, Google, read Nov 11, 2026
• Google Gemini Decision 1 JUST LAUNCHED! Tested vs Liquid d1-ultra & Jev!, AI WITH Rithesh, Nov 12, 2026