Google Releases Gemini Decision 1 Decision Model and Decisions API
Google has released the Gemini Decision 1 model and the Gemini Decisions API powered by it. The model does not generate token by token; instead, it scores preset candidate answers through a single forward pass, supporting three types of tasks: yes/no judgments, multiple choice with up to 255 options, and scoring according to scoring criteria. A single request can process up to 1,000 questions, supports text, image, audio, video, and PDF inputs, has a context window of up to 256K tokens, end-to-end latency of approximately 150–250 milliseconds, and charges $0.10 per million input tokens with no charge for output tokens. Google says it ranked first in 7 out of 9 benchmarks, with 92.2% accuracy on JevBench, but it trails TypeSafe Jev 1.14 on When2Call and Mind2Web Multimodal; internally it has already been used for model routing in Gemini Notebook, reducing latency for everyday queries by up to 40%.
Why read this
AI score 75
Google launches a decision model that does not generate token by token but directly scores candidate answers, and provides benchmark comparisons with general-purpose models and specialized decision models.
Topics
Body · AI translation
While the entire AI industry is still competing over who can generate more tokens and whose reasoning capabilities are stronger, Google has instead launched a new model that does not rely on generating answers word by word.
Just now, Google officially released Gemini Decision 1, along with the Gemini Decisions API powered by this model. Unlike traditional generative large models, this model is specifically designed for rapid decision-making. It does not need to generate text token by token; instead, through a single forward pass, it scores a pre-defined set of candidate answers.
Google CEO Sundar Pichai also posted on X to introduce the main capabilities of Gemini Decision 1. He wrote:
Today, we are officially launching the Gemini Decisions API in Google AI Studio, powered by the all-new Gemini Decision 1 model.
Unlike traditional models that generate text token by token, this model is specifically designed for decision tasks with a closed set of options. In a single processing pass, it can score up to 1,000 questions across text, images, audio, video, and PDFs.
Here are a few noteworthy details:
· Supports three decision types: yes/no judgments, multiple choice (up to 255 options), and rubric-based scoring;
· Priced at only $0.10 per million input tokens, with zero cost for output tokens, about one-fifth the cost of Flash-Lite;
· Leads in multiple benchmark evaluations while also having the lowest calibration error.
Teams inside Google have already begun using this technology.
In the past, when developers wanted AI to determine whether an email was spam, or to decide which Agent a task should be assigned to, they often needed to call a generative large model and then extract the result from the text generated by the model. Gemini Decision 1 eliminates the step of token-by-token decoding.
Developers only need to provide the input content, the question to be judged, and the candidate options, and the model can directly return the corresponding judgment result, probability, and confidence.
According to figures published by Google, Gemini Decision 1 can process up to 1,000 independent questions in a single request, supports multimodal inputs such as text, images, audio, video, and PDFs, and has an end-to-end latency of only 150–250 milliseconds.
Pricing is also fairly aggressive: only $0.10 per million input tokens, with no charge for output tokens. In the same set of benchmarks published by Google, its invocation cost is nearly 5 times lower than that of Gemini 3.5 Flash-Lite.
More notably, this is not just a model for answering multiple-choice questions. Google has already used it in scenarios such as AI model routing, Agent tool selection, product design, and safety detection, attempting to make it a foundational component specifically responsible for judgment and decision-making in AI systems.
Processing 1,000 questions at once, with up to 255 options
According to the technical introduction published by Google, Gemini Decision 1 provides three core decision-making capabilities.
The first is Yes/No (binary judgment). For example, determining whether there is a safety hazard in an image, or whether a piece of text violates established rules. The model gives the probability that the corresponding condition holds.
The second is Multiple Choice. Developers can preset 2–255 candidate options, and the model not only selects the answer that best fits the conditions but also returns the normalized probability for each candidate option.
The third is Ordered Rubric. Developers can define scoring criteria at different levels themselves, letting the model evaluate the input content and return a continuous weighted score along with the probability of each level.
The three types of tasks can also be executed in a mixed manner within a single request.
Google provided a simple example: upload a photo of an office break room counter to the model, along with a text description, and ask it to simultaneously determine whether there is any liquid spill on the countertop, what cleanup measures should be taken, and the severity of the situation.
The traditional approach might require designing multiple classifiers, or using a generative large model to execute different prompt tasks. Gemini Decision 1 can complete these judgments in a single request.
In addition, the model supports a multimodal input context of up to 256K tokens. For enterprise scenarios, this means developers can analyze call recordings together with customer materials, or make judgments by combining insurance claim documents with on-site photos.
Each decision provides a confidence score from 0 to 1, making it easy for the system to hand uncertain results over to other models or human review.
Another competitive advantage of Gemini Decision 1 is speed and cost.
According to Google's official introduction, in benchmark evaluations covering Agent routing, reasoning, ranking, and multimodal understanding, Gemini Decision 1 performs leadingly and achieves the lowest calibration error.
In test contexts of no more than 16K tokens, the median latency of its model computation stage stays at 50–100 milliseconds, with end-to-end latency of about 150–250 milliseconds. Even when the number of questions in a single request increases from 1 to 100, latency remains essentially stable.
On the common benchmark test sets used officially, the cost of the Gemini Decisions API is nearly 5 times lower than calling Gemini 3.5 Flash-Lite through the Interactions API.
First place in 7 of 9 tests, with the highest accuracy at 92.2%
With this release of Gemini Decision 1, Google also published a rather noteworthy model comparison scorecard.
Unlike the usual approach of comparing general-purpose large models together, Google this time selected three competing products: OpenAI's GPT-6 Luna, TypeSafe Jev 1.14, which is specifically aimed at classification decisions, and its own Gemini 3.5 Flash-Lite.
Judging from the results, Gemini Decision 1 took first place in 7 of 9 benchmark tests and also leads in probability calibration, but in the other two tests and in total invocation cost, there are still cases where competitors perform better.
In the JevBench test, Gemini Decision 1 achieved 92.2% accuracy, surpassing GPT-6 Luna's 90.9%, TypeSafe Jev 1.14's 90.0%, and Gemini 3.5 Flash-Lite's 86.6%.
In the Amazon ESCI product search relevance test, Gemini Decision 1 achieved a Macro-F1 score of 58.7%, 4 percentage points higher than GPT-6 Luna and 11.5 percentage points higher than Flash-Lite.
In the New Yorker Caption test, the gap widened further. Gemini Decision 1 achieved 79.7%, while GPT-6 Luna scored 68.4%, a difference of 11.3 percentage points. This test involves understanding and selecting cartoon captions, testing the model's contextual understanding and ability to judge candidate content.
In the medical question-answering test PubMedQA, Gemini Decision 1 also ranked first with a score of 79.8%, higher than GPT-6 Luna's 74.4%.
The results of the multilingual test Global-MMLU-Lite are similar. Gemini Decision 1 achieved 86.0%, while GPT-6 Luna scored 79.6%, a difference of 6.4 percentage points.
However, Google did not achieve across-the-board leadership.
In the Mind2Web Multimodal test, TypeSafe Jev 1.14 ranked first with 86.8%, while Gemini Decision 1 scored 86.2%, trailing by 0.6 percentage points.
In the When2Call test, which examines the ability to judge tool invocation, TypeSafe Jev 1.14's advantage was even more obvious, scoring 82.0% versus Gemini Decision 1's 79.9%.
This result is particularly worth noting. Tool selection and judgment of invocation timing are among the Agent application scenarios that Gemini Decision 1 focuses on, yet in this test the specialized decision model TypeSafe Jev still maintained its lead.
In addition, in the audio understanding test MMAU and the video understanding test Video-MME, Gemini Decision 1 achieved 71.8% and 69.1%, respectively, both higher than Flash-Lite.
More noteworthy than accuracy is the probability calibration capability that Google particularly emphasized this time.
In the Top-1 ECE (Expected Calibration Error) test, Gemini Decision 1's calibration error was only 0.0250, while GPT-6 Luna's was 0.0721, TypeSafe Jev 1.14's was 0.0494, and Flash-Lite's reached 0.2104.
In other words, in this set of evaluations, Gemini Decision 1's calibration error is about 65% lower than GPT-6 Luna's and about 88% lower than Flash-Lite's.
This metric is especially important for Agent systems.
For example, when an Agent needs to determine whether a piece of code has a security vulnerability, the model gives not only a result of "exists" or "does not exist," but may also include a probability value. The system can decide, based on the relationship between the probability and the actual accuracy rate, whether to take action automatically or hand it over to a human for further verification.
Google has already put it to use: Agent routing latency reduced by up to 40%
Unlike some new models that only showcase benchmark results, Google this time also disclosed application cases of the Gemini Decisions API in internal products.
The most noteworthy of these is Gemini Notebook.
Google is using the Decisions API as a real-time model router to select the appropriate model tier and thinking intensity based on the complexity of user requests. For simple everyday questions, the system can prioritize lower-cost, faster-responding models; for complex reasoning tasks, it then hands the request to more capable models.
Google says that with this dynamic routing approach, latency for everyday queries can be reduced by up to 40%.
This design is closely related to the current development trend of Agent systems. When an Agent completes a task, it often needs to continuously judge the next action, such as which tool to choose, which model to call, or whether human intervention is needed.
If every simple decision had to be handled by a large generative model, the system would incur additional latency and cost. A specialized decision model offers another option: hand relatively simple judgments with clear options to a low-latency model, and leave complex tasks to general-purpose large models.
In addition to Gemini Notebook, Google also disclosed three other application directions.
In Google Stitch, the newly launched Rapid Canvas experimental feature uses the Decisions API to form interactive product flow wireframes within seconds and quickly respond to modification instructions issued by users via voice or clicks. After users confirm the design direction, the page can be transferred to the full Stitch workflow for further refinement.
In Google Labs' CC product, the Decisions API is used to quickly select the appropriate emoji reaction from more than 200 emoji options. Within Google Search, Google also uses this API to build a security detection layer for MCP tools, identifying potential adversarial behavior and content.
Reference links:
This article comes from the WeChat public account "AI Frontline" (ID: ai-front), author: Dongmei, published by 36Kr with authorization.