Hopp til hovedinnhold

Google launches Gemini Decision 1: the AI model that makes 1,000 choices in a quarter of a second

The new model does not write answers, it chooses: yes/no, multiple choice and rubric scores on up to 1,000 questions in one call. Output is free, but Google does not say where data is processed.

Av Joachim HøgbyKilde: Google

On Wednesday evening Google launched the Gemini Decisions API, a new developer interface that does not write answers but makes choices. Behind it sits a dedicated model, Gemini Decision 1. It answers yes/no questions, picks between fixed options and grades against a rubric, and it can handle up to 1,000 such questions in a single call.

Google CEO Sundar Pichai announced the launch on X on Wednesday 11 November at 19:15 Norwegian time (18:15 UTC), alongside a post on Google's blog. The API is available from the same day through the Gemini API in Google AI Studio.

A model that chooses instead of writing

Ordinary language models generate their answer token by token. That takes time, and the customer pays for every token the model writes. Gemini Decision 1 works differently. According to Google, the model scores all candidate answers in a single forward pass and returns probabilities and a confidence score between 0 and 1 for each answer.

Google describes three question types:

  • Yes/no: the probability that a condition holds.
  • Multiple choice: from 2 to 255 options, with the winner and a probability for every option.
  • Rubric score: a grade against an ordered rubric, with a weighted score.

Input can be text, images, audio, video and PDFs, also combined, up to 256,000 tokens. Google gives examples such as a call recording paired with account data, or a scanned claim form paired with photos of the damage.

Answers in a quarter of a second, billing only for input

Google states 150–250 milliseconds end to end, and 50–100 milliseconds of model compute for contexts up to 16,000 tokens. According to the company, latency stays nearly flat from 1 to 100 questions per request.

The price is $0.10 per million input tokens. Output is free, and the shared input is billed once per request no matter how many questions are asked. There is also a no-cost tier for prototyping in Google AI Studio.

Google claims Gemini Decision 1 leads on the company's own benchmarks for agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. On the company's shared benchmark suite it is said to be nearly five times cheaper than Gemini 3.5 Flash-Lite. These are Google's own measurements. Independent tests have not been published.

Google already uses it internally

According to Google, several of its own products already use the API:

  • Gemini Notebook routes each request to the right model tier. Google cites up to 40 percent lower latency on everyday queries.
  • Google Search uses it as a safety layer for MCP tools, to detect adversarial behaviour and content.
  • Stitch and CC from Google Labs use it for rapid design edits and to pick the right emoji from more than 200.

Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Support in Gemini in Google Cloud is coming "soon", according to Google, with no date given.

What it means for businesses

Many AI deployments in companies are in practice classification: sorting email, checking documents against rules, picking the right case handler or deciding whether a case needs a human. Many teams use a general language model for this today, and pay for text nobody reads.

  • Cost and speed. If you have large volumes of simple yes/no or category decisions, a model like this could cut both latency and the bill. Test on your own data before trusting Google's figures.
  • Human in the loop. The confidence score makes it possible to route uncertain cases to a person. That matters where mistakes are costly, such as insurance, customer service and compliance.
  • Gaps in the announcement. Google does not say where data is processed, whether it can be restricted to the EU, or which terms apply to enterprise customers. Enterprise support via Google Cloud is not yet available. Clarify this before sending personal data or sensitive documents.

Sources and media

Illustration: AI-generated with Grok Bot / hogby.ai

Mer om AI

Oppskriften. Strategien. Hver dag.

AI-nyheter for ledere. Kuratert av en CIO som bygger det selv. Daglig i innboksen.