A model and API that return only decisions
Google has released "Gemini Decision" (gemini-decision-1), a model that returns probabilities for yes/no judgments, choices among options and graded scoring, along with its dedicated "Decisions API." The changelog for its developer-facing "Gemini API" marks it "Released" in an entry dated Nov. 11 U.S. time. In the official model list, it appears in the section for models designed for specific tasks.
The list describes it as "Fast, calibrated multimodal yes/no, choice, and score decisions with zero output tokens." Because it does not generate text and returns answers as probabilities, output tokens are always zero. The model card describes it as a dedicated model built for the Decisions API.
As uses, the company cites routing inquiries, checking compliance with many rules at once, scoring candidates against evaluation criteria, selecting tools for AI agents and low-latency classification of streaming video and audio. It said the approach reduces latency and parsing work compared with generating text and then parsing it.
@gptbotio, a bot that detects API changes, reported it on X at 1:45 a.m. Nov. 12 Japan time.
Proposed changes to 4 SDKs merged early on Nov. 12
Shortly after midnight on Nov. 12, ORIENS reported that proposed changes (pull requests) adding client.decisions.create, which calls the Decisions API, had been filed for four of Google's official SDKs (software development kits): Python, JavaScript, Go and Java. At that point, the proposed changes had not been merged, and the official documentation did not mention the API.
All four proposed changes were merged at 2:11 to 2:12 a.m. Nov. 12 Japan time. The official guide says SDK version 3.2 or later is required to use the Decisions API.
Three types of questions, up to 1,000 per request
According to the official guide, a request contains three things: the model to use, shared input that serves as material for the decisions, and a list of questions. There are three types of questions: noul, which asks yes or no; choice, which selects one of a fixed set of options; and score, which grades against graded evaluation criteria. A single request can contain 1 to 1,000 questions. Questions cannot see one another and are judged together against the shared input in a single pass. The company said latency barely changes as more questions are added.
Input can be text, images, audio, video or PDFs. The input limit is 262,144 tokens, including the shared input and all questions. Options for choice and levels for score can each be set from 2 to 255.
Each answer comes with "confidence," an estimate of the probability that the most likely answer is correct. The company shows a use in which this value determines whether processing proceeds automatically or is sent for human review. The model is billed as "calibrated," but the model card and guide do not include accuracy evaluation results.
| Type | Question | Returned values |
|---|---|---|
| noul | Yes or no | Probability of "yes" (0.01 to 0.99), confidence |
| choice | One of 2 to 255 options | Most probable option, probability of each option, confidence |
| score | Graded against criteria with 2 to 255 levels | Probability-weighted score, probability of each level, confidence |
Paid tier costs $0.10 per million input tokens, with no charge for output
According to the price table, the paid tier costs $0.10 per million input tokens, the same for text, images, video, audio and PDFs. Output is zero tokens and therefore not billed, and on the free tier input is also free. Shared input is billed only once regardless of the number of questions, and each question adds tokens for its instructions and criteria.
Input and output on the free tier are used to improve Google's products, while those on the paid tier are not. The model does not support the "Batch API," which processes large volumes of requests cheaply at once, or "context caching," which lowers costs by reusing the same input. It also does not support "Flex" and "Priority" inference, which change processing priority, and when the same input is sent across requests, the full amount is counted each time.
Usage limits (rate limits) on the free tier are 200 requests per minute and 10,000 per day. On the highest paid tier (Tier 3), the limit is 15,000 requests per minute, with no daily cap.
| Usage tier | Requests per minute | Tokens per minute | Requests per day |
|---|---|---|---|
| Free | 200 | 50,000 | 10,000 |
| Tier 1 | 1,500 | 1,000,000 | 50,000 |
| Tier 2 | 5,000 | 4,000,000 | No limit |
| Tier 3 | 15,000 | 10,000,000 | No limit |
Offered only on v1beta of the Gemini Developer API
The official guide says the Decisions API is available on v1beta of the Gemini Developer API and that Gemini Decision is the only supported model. The changelog explicitly says "generally available (GA)" in the previous day's Nov. 10 entry for Gemini 4 Carbon, but the Gemini Decision entry says only "Released."
The official pages read do not mention availability outside the Gemini Developer API, such as on Google Cloud's Vertex AI. ORIENS has not run tests with an API key and has not confirmed whether it can actually be called.
Update, Nov. 12: Google published an official blog post at 3 a.m. Nov. 12 Japan time explaining Gemini Decision and the Decisions API in detail. It said the time to return an answer is 150 to 250 milliseconds end to end. On a common set of benchmarks, using Gemini Decision through the Decisions API is nearly 5 times cheaper than using Gemini 3.5 Flash-Lite through the Interactions API, Google said. Within Google, teams including Gemini Notebook and Google Search are already using the Decisions API, it said. Support for external developer tools such as OpenRouter and LangChain is being expanded in stages, and availability in Gemini on Google Cloud is coming soon, the company said.
