Google's new Gemini Decision 1 answers yes/no without writing a word. The builder tip: stop paying a chat model for routing

• Yes/no (Google calls it noul): the probability that a condition holds.
• Multiple choice: 2 to 255 options, with a winner and a probability for each.
• Rubric score: an ordered scale starting at 0, with a weighted score and per-level probabilities.

Every answer comes with a confidence score from 0 to 1.

Price. Per Google's pricing page (opens in a new tab): $0.10 per 1M input tokens on the paid tier, and no output tokens at all. The shared input is billed once per request, however many questions you ask; each question adds only its own instruction tokens. There is a free tier in AI Studio for prototyping.

Google's own claims (not checked by MODEL). 150-250 ms end to end; latency nearly flat from 1 to 100 questions per request; the lowest calibration error on Google's benchmark suite; and "nearly 5x cheaper" across that suite than Gemini 3.5 Flash-Lite through the Interactions API. Google says it already uses the API inside Gemini Notebook for model routing, in Stitch, and as a safety layer for MCP tools in Search.

Where else. Vercel AI Gateway (opens in a new tab) and OpenRouter (opens in a new tab) listed it the same day. Google says LangChain, LiteLLM, Pydantic AI and others are rolling out, with Gemini in Google Cloud "coming soon."

The builder tip

Look at your agent's logs and count the calls where a big chat model only had to pick: which tool, which model tier, is this allowed, is this a refund request. Each one pays for generated text you parse and throw away.

• Split the decision from the work. Send routing, policy checks and tool picks to a scoring endpoint. Keep the generative model for the actual answer.
• Batch the questions. Ask all your checks against the same input in one request. The input is billed once.
• Use the confidence score. Pick a threshold. Below it, fall back to a bigger model or a human. Don't let a 0.55 "yes" approve a payment.
• Test on your own data first. Run a few hundred of your past decisions through it and compare with what actually happened before you switch.

The catch. Google's examples call it through the v1beta API version, which Google's docs describe as early features that may change. On the free tier, Google's pricing page says your content is used to improve its products; on the paid tier it is not. The Batch API and context caching are not supported. And the speed, calibration and cost numbers are Google's own.

Disclosure: Not tested by MODEL. Not sponsored. No affiliate links. Facts from Google's Nov. 11, 2026 announcement and pricing page, plus Vercel and OpenRouter listings.

MODEL Editors