Tuesday, November 17, 2026

News

Google launches Gemini Decisions API: model returns confidence scores instead of text, $0.10 per million tokens

ModelsPatryk Raba
Google launches Gemini Decisions API: model returns confidence scores instead of text, $0.10 per million tokens
Fot. Asoundd, Wikimedia Commons (CC BY-SA 4.0)

Google has added a new endpoint to the Gemini API and AI Studio, built on the Gemini Decision 1 model. Instead of generating text, it scores inputs against a list of questions and returns probabilities and confidence values, with Google citing 150-250 ms end-to-end latency for contexts up to 16K tokens.

Contents
  1. How it works
  2. Pricing and speed
  3. Benchmarks and use cases
  4. Availability and what's next

Google DeepMind has introduced the Gemini Decisions API, a new endpoint in the Gemini API and Google AI Studio that does not write answers word by word. The Gemini Decision 1 model scores the input against a set of questions and returns probabilities along with a confidence score. It is meant for fast decisions in applications and agents.

How it works

According to a post by Min Choi and Felix Yu of Google DeepMind, the endpoint is designed for closed-set decisions: tool selection, policy checks, model routing or grading the usefulness of context. The model evaluates the options in a single pass and returns probabilities and confidence values on a scale from 0 to 1.

There are three question types. A yes/no question returns a probability, with optional criteria describing true and false. A choice question covers 2 to 255 options and returns normalized probabilities for each. A score question grades the input against an ordered scale starting at 0 and returns a weighted score and probabilities for each level.

The shared input can combine different types of material, and the resolution of images and video can be tuned to balance visual detail against token cost. The data format is the same as in the Interactions API. In the post's example, the model analyzes a photo of a breakroom counter and a text note, answering questions of all three types.

Pricing and speed

Google bills only input tokens: $0.10 per million, with no charge for output tokens or thinking tokens, and no separate fees for audio transcription or OCR. The shared input is counted once per request, and each question adds only the tokens of its own instruction and criteria. AI Studio also offers a no-cost tier for prototyping.

The company says that for multimodal inputs and contexts up to 16K tokens, latency is 150-250 ms end-to-end and stays nearly flat as the number of questions grows from 1 to 100. Google also claims the solution is nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. That comparison comes from the vendor.

Benchmarks and use cases

Google says Gemini Decision 1 leads across a benchmark suite spanning agent routing, reasoning, ranking and multimodal understanding, and has the lowest calibration error. The post contains only charts, with no figures in the text, so the results cannot be independently verified here.

Among internal uses, Google lists Gemini Notebook, where the model routes queries to the appropriate model and thinking level, which is meant to deliver up to 40% lower latency on everyday queries. Google Search uses it as a safety layer for MCP tools, detecting adversarial behavior and content. The Google Labs project CC picks an emoji reaction from more than 200 options.

Availability and what's next

The endpoint is available in the Gemini API and Google AI Studio, with guides for Python, JavaScript and REST. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, as well as the google-genai SDKs. Support in Gemini in Google Cloud is announced as coming soon.

For teams building agents, it offers a way to offload simple steps such as routing or filtering from large models. Because every answer carries a confidence score, uncertain cases can be sent to a human or a fallback model. How well it performs in a given use case still has to be tested on your own data, since the quality claims come from Google.

Share: