Intel / blog

Gemini Decisions API: Fast, calibrated multimodal decisions

Source: blog.google
Date: Published Nov 11, 2026

From the Daily Brief

Google DeepMind introduced the Gemini Decisions API and its first model, gemini-decision-1. Instead of generating text, it answers yes/no, multiple-choice and rubric questions, up to 1,000 of them in a single call, and returns a probability for each answer. Inputs can be text, images, audio, video or PDFs. Simon Willison tested it the same day and reports that it costs 10 cents per million input tokens with no charge for output, and that its answers come back as graded probabilities rather than hard 0 or 1 results. He compares it with Jev and with OpenAI's equivalent, and notes that Jev does not accept video, audio, images or PDFs. Google's python-genai 3.2.0 release already exposes the endpoint as client.decisions.create. The practical shift is for steps such as routing a request, grading an output or checking a policy: those decisions now run as one scoring call with a confidence value attached, which is cheaper and faster than prompting a general model once per question and parsing its prose.

Read the 2026-11-12 Brief →

Context

Google's new Gemini Decisions API answers questions that have a fixed set of possible answers, and it does so without writing any text (Google blog). A developer sends one input, which can mix text, images, audio, video and PDFs up to 256,000 tokens, along with as many as 1,000 questions about it. Each question is a yes-or-no check, a pick from 2 to 255 supplied options, or a grade on an ordered scale. The model behind it, Gemini Decision 1, scores the candidate answers directly and returns a probability for each one, plus a confidence score from 0 to 1, instead of generating a reply word by word.

That design is meant for the small judgments inside apps and agents, such as choosing which model or tool should handle a request, checking content against a policy, or grading an input. Because the result is a probability rather than prose, a program can act on a confident answer and send an uncertain one to a fallback or a person. Google gives the example of checking a photo of a break room counter and asking at once whether there is a spill, which cleanup action fits, and how severe the mess is.

Google's case is speed and cost. It says median responses take 150 to 250 milliseconds end to end for inputs up to 16,000 tokens, and stay nearly flat from 1 to 100 questions per request. Pricing is $0.10 per million input tokens with no output charge, and the shared input is billed once however many questions are asked. Google reports that Gemini Decision 1 leads a set of benchmarks covering routing, reasoning, ranking and understanding, with the lowest error, meaning its stated confidence most closely matches how often it is right, but the post's text does not name the benchmarks or the models it was compared with. Google's own products use it, including Gemini Notebook, which routes queries to different model tiers. OpenAI launched an API with the same name and a similar three-question design in October (OpenAI guide, Intel). OpenAI's guide lists text and images as inputs and says audio and files are not supported, and the Google post does not compare the two.

Key takeaways · AI-distilled

Google says Gemini Decision 1 scores candidate options in one forward pass instead of generating text. It reports 150-250 ms median end to end, nearly flat from 1 to 100 questions per request, for contexts up to 16K tokens.
Pricing is $0.10 per 1M input tokens with no output-token charge, and the shared input is billed once per request however many questions it carries. Google claims that makes it nearly 5x cheaper than Gemini 3.5 Flash-Lite on its benchmark suite.
A request takes one multimodal input of up to 256K tokens (text, images, audio, video, PDFs) and up to 1,000 questions of three kinds: yes/no, choice among 2 to 255 options, and ordered rubric scores. Every answer carries a 0-1 confidence score.
Google reports internal use: Gemini Notebook routes queries to model tiers with it, serving everyday queries with up to 40% lower latency, and Google Search uses it to detect adversarial behavior around tools.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.

Read the source blog.google