Google Launches Gemini Decisions API for Batched Judgment Calls
Google launched the Decisions API in Gemini API alongside the Gemini Decision 1 model, which skips text generation and returns structured yes/no, multiple-choice and scoring judgments with calibrated confidence for up to 1,000 questions per request. Pricing is $0.10 per million input tokens with no output charge, latency is 150-250 ms for inputs under 16K tokens, and Google claims it is nearly 5x cheaper than generating answers with Gemini 3.5 Flash-Lite. The API turns routing, moderation and grading calls from generate-then-parse into direct calibrated probabilities usable as automation thresholds.
Full Report
宝玉 (@dotey) · 2026-11-11 22
Google launches the Gemini Decisions API: it doesn't write answers, it only makes judgments, you can ask up to 1,000 questions at once, $0.1 per million input tokens, and output is free
Google today launched the Decisions API within the Gemini API, along with a new companion model, Gemini Decision 1. It doesn't generate text; it only answers questions: give it an input and a set of questions, and it directly returns the answer to each question, plus a confidence score for "how sure it is."
[What problem does it solve]
When building AI applications, many calls really just need a single judgment: which team should this customer service message be routed to? Should this question go to a cheap model or a reasoning model? Does this image violate policy? Previously you could only have a large model write a paragraph and then parse the answer out of it—slow, prone to parsing failures, and you still had to pay for output tokens.
The Decisions API has only three question types: yes/no questions, which return the probability of "yes"; multiple-choice questions, with 2 to 255 options, which return the selected option and the probability of each option; and scoring questions, which score according to the tiers you define. Each answer comes with a confidence score from 0 to 1, and code can set thresholds—for example, execute automatically above 0.7, and hand off to a human below that. Google says these probabilities are calibrated, meaning that when the model says 90%, it is actually right about nine times out of ten, so the confidence can be used directly as a threshold.
An example from the Google blog: upload a photo of a break room countertop and ask three questions at the same time. "Is there any spilled liquid" answers 0.95; "How to handle it" selects "wipe it with a cloth," with confidence 0.98; "Rate the messiness from 0 to 2" gives 1.19, with confidence 0.77.
[How fast, how cheap]
A single request can take up to 256K input tokens, and text, JSON, images, audio, video, and PDF can be mixed together—for example, a recording of a customer service call plus account information. Up to 1,000 questions can be attached to the same input, and they don't affect each other. Input is billed only once; each additional question only adds the word count of that question itself.
Google claims that when the input is within 16,000 tokens, end-to-end latency is 150 to 250 milliseconds, and asking 1 question takes about the same time as asking 100; under the same batch of evaluations, it is nearly 5 times cheaper than generating answers with Gemini 3.5 Flash-Lite. Gemini Notebook already uses it to pick a model for each query, cutting latency for everyday questions by up to 40%.
[Who can use it]
You can call it with an API key from Google AI Studio, using the model name gemini-decision-1 through the v1beta interface. There is a free tier of 200 requests per minute and 10,000 per day; data from the free tier will be used by Google to improve its products, while the paid tier's data will not. OpenRouter, LangChain, Vercel, and others are integrating it, and Gemini on Google Cloud will support it a bit later.
Context caching and the Batch API are not supported, so if you ask repeatedly about the same large document, input is billed at full price each time; you cannot pass YouTube links directly; and the questions themselves can only be written in text.
[Compared with OpenAI's product of the same name]
OpenAI built the same thing a month earlier, also called the Decisions API: a beta on October 6, officially opened on October 22, using GPT-6 Luna, also at $0.1 per million input tokens with no charge for output. The differences are that OpenAI allows up to 400 questions at a time and only accepts text, images, and JSON—not audio, video, or PDF; but it supports caching, so repeated input gets a discount, and it can be inserted directly into conversation history and tool-call records.
The startup TypeSafe's Jev 1.14 also specializes in this kind of judgment, at $0.042 per million input tokens on OpenRouter, accepting only text and images.
In the comparison table attached to the Google blog, Gemini Decision 1 ranks first in 7 of 9 tests (among them, the audio and video tests had only it and Flash-Lite participating), with a calibration error of 0.025 versus 0.072 for GPT-6 Luna; it lost to Jev 1.14 on web operation and "should a tool be called." For the total cost of the 7 shared evaluations, Jev was the lowest at $0.68, Gemini Decision 1 was $0.88, and GPT-6 Luna was $0.93. All of the above benchmark results were measured by Google itself.
Google blog: https://t.co/P5hlLj7fXN Official documentation: https://t.co/WphCUZv5oT OpenAI Decisions API documentation: https://t.co/MGgS32OuNy
↳ Quoting @sundarpichai (Sundar Pichai): Today we're launching the Gemini Decisions API, powered by our new Gemini Decision 1 model in @GoogleAIStudio.
Instead of generating text token by token, it's built for closed-set choices scoring up to 1,000 questions across text, images, audio, video + PDFs in a single pass. A few details:
- 3 decision types: yes/no, multiple choice (up to 255 options) + rubric scoring
- 0 billed for output tokens (~5x cheaper than Flash-Lite)
- leads on several benchmark evals with lowest calibration error
Google teams are already building with it, below is a look at Stitch Rapid designing a full product flow in seconds (1x speed)
