Google Gemini Decision 1 released: does not generate verbatim, judges 1000 questions in one request

Author | Dongmei (AI Frontline)

While the entire AI industry is still competing on who can generate more tokens and whose reasoning ability is stronger, Google has launched a new model that does not rely on generating answers word by word.

Recently, Google officially released Gemini Decision 1 and simultaneously launched the Gemini Decisions API powered by this model. Unlike traditional generative large models, this model is specifically designed for rapid decision-making, scoring pre-defined candidate answers through a single forward pass without generating text token by token.

Google CEO Sundar Pichai also introduced the main capabilities of Gemini Decision 1 on the X platform. He wrote:

Today, we officially launched the Gemini Decisions API in Google AI Studio, supported by the brand-new Gemini Decision 1 model.

Unlike traditional models that generate text token by token, this model is designed for decision-making tasks within a closed set of options. It can score up to 1000 questions in a single processing session across text, images, audio, video, and PDFs.

Here are some noteworthy details:

• Supports three types of decision-making: Yes/No judgments, multiple choice (up to 255 options), and scoring based on criteria;
• The price is only $0.10 per million input tokens, with no cost for output tokens, making it about one-fifth the cost of Flash-Lite;
• It leads in multiple benchmark evaluations while having the lowest calibration error.

Google's internal teams have already begun using this technology.

In the past, when developers wanted AI to determine whether an email was spam or which agent a task should be assigned to, they often needed to call a generative large model and extract results from the generated text. Gemini Decision 1 eliminates the need for word-by-word decoding.

Developers only need to provide input content, the question to be judged, and candidate options, and the model can directly return the corresponding judgment result, probability, and confidence.

This means that for tasks that only require a classification, scoring, or selection result, AI no longer needs to complete a full text generation process.

According to data released by Google, Gemini Decision 1 can handle up to 1000 independent questions in a single request, supporting multimodal inputs such as text, images, audio, video, and PDFs, with an end-to-end latency of only 150 to 250 milliseconds.

The pricing is also quite aggressive: only $0.10 per million input tokens, with no charge for output tokens. In the same set of benchmark tests released by Google, its calling cost is nearly five 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 applied it in scenarios such as AI model routing, agent tool selection, product design, and security detection, aiming to make it a foundational component responsible for judgment and decision-making in AI systems.

Processing 1000 Questions at Once, Up to 255 Options

According to the technical introduction released by Google, Gemini Decision 1 provides three core decision-making capabilities.

The first type 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 will provide the probability of the corresponding condition being true.

The second type is Multiple Choice. Developers can preset 2 to 255 candidate options, and the model will not only select the most suitable answer but also return the normalized probability for each candidate option.

The third type is Ordered Rubric (ordered standard scoring). Developers can define different levels of scoring criteria, allowing the model to evaluate the input content and return continuous weighted scores and probabilities for each level.

These three tasks can also be mixed in a single request.

Google provided a simple example: upload a photo of an office kitchenette countertop to the model along with a text description, asking it to simultaneously determine whether there is liquid spillage on the countertop, what cleaning measures should be taken, and the severity of the situation.

Traditional methods might require designing multiple classifiers or executing different prompt tasks through generative large models. Gemini Decision 1 can complete these judgments in a single request.

Additionally, the model supports multimodal input contexts of up to 256K tokens. For enterprise scenarios, this means developers can analyze call recordings alongside customer profiles or combine insurance claim documents with on-site photos for judgment.

Each decision will provide a confidence level between 0 and 1, making it easy for the system to pass uncertain results to other models or human review.

Another competitive edge of Gemini Decision 1 is its speed and cost.

According to official information from Google, in benchmark evaluations covering agent routing, reasoning, ranking, and multimodal understanding, Gemini Decision 1 performed leadingly and achieved the lowest calibration error.

In test contexts not exceeding 16K tokens, the median latency during the model computation phase remains at 50 to 100 milliseconds, with an end-to-end latency of about 150 to 250 milliseconds. Even when the number of questions in a single request increases from 1 to 100, the latency remains relatively stable.

In the common benchmark test set used by officials, the cost of Gemini Decisions API is nearly five times lower than calling Gemini 3.5 Flash-Lite through the Interactions API.

9 Tests, 7 First Places, Highest Accuracy of 92.2%

In this release of Gemini Decision 1, Google also published a compelling model comparison scorecard.

Unlike the usual practice of comparing general large models, Google chose three competing products this time: OpenAI's GPT-6 Luna, TypeSafe Jev 1.14 specifically aimed at classification decisions, and its own Gemini 3.5 Flash-Lite.

From the results, Gemini Decision 1 secured first place in 7 out of 9 benchmark tests, also leading in probability calibration, but there were still cases where competitors performed better in the other two tests and overall calling costs.

In the JevBench test, Gemini Decision 1 achieved an accuracy of 92.2%, 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 scored 58.7% in Macro-F1, 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 scored 79.7%, while GPT-6 Luna scored 68.4%, a difference of 11.3 percentage points. This test involved understanding and selecting captions for cartoons, testing the model's contextual understanding and candidate content judgment ability.

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 were similar. Gemini Decision 1 scored 86.0%, while GPT-6 Luna scored 79.6%, a difference of 6.4 percentage points.

However, Google did not achieve a comprehensive lead.

In the Mind2Web Multimodal test, TypeSafe Jev 1.14 ranked first with a score of 86.8%, while Gemini Decision 1 scored 86.2%, trailing by 0.6 percentage points.

In the When2Call test, which assesses tool invocation judgment ability, TypeSafe Jev 1.14's advantage was even more pronounced, scoring 82.0%, while Gemini Decision 1 scored 79.9%.

This result is particularly noteworthy. Tool selection and invocation timing judgment is one of the key application scenarios for Gemini Decision 1, but in this test, the specialized decision model TypeSafe Jev still maintained its lead.

Additionally, in the audio understanding test MMAU and the video understanding test Video-MME, Gemini Decision 1 scored 71.8% and 69.1%, respectively, both higher than Flash-Lite.

More noteworthy than accuracy is the probability calibration ability 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 was 0.0721, TypeSafe Jev 1.14 was 0.0494, and Flash-Lite reached 0.2104.

In other words, in this set of evaluations, the calibration error of Gemini Decision 1 is about 65% lower than that of GPT-6 Luna and about 88% lower than that of Flash-Lite.

This metric is particularly important for Agent systems.

For example, when an Agent needs to determine whether a piece of code has security vulnerabilities, the model provides not only a result of "exists" or "does not exist," but may also include a probability value. The system can decide whether to take automatic action or hand it over to humans for further verification based on the relationship between the probability and the actual accuracy.

Google has implemented it: Agent routing delay reduced by up to 40%

Unlike some new models that only showcase benchmark scores, Google has also disclosed application cases of the Gemini Decisions API in internal products.

The most noteworthy is Gemini Notebook.

Google is using the Decisions API as a real-time model router, selecting the appropriate model tier and thinking intensity based on the complexity of user requests. For everyday simple questions, the system can prioritize calling lower-cost, faster-responding models; for complex reasoning tasks, it can then hand the request over to more capable models.

Google claims that with this dynamic routing approach, the delay 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 has to be handled by a large generative model, the system would incur additional delays and costs. A dedicated decision model provides an alternative: handling relatively simple, clearly defined judgments with low-latency models while leaving complex tasks to general large models.

In addition to Gemini Notebook, Google has also disclosed three other application directions.

In Google Stitch, the newly launched Rapid Canvas experimental feature utilizes the Decisions API to create interactive product flow wireframes in seconds and quickly respond to user modification instructions issued through voice or clicks. Once users confirm the design direction, they can transfer the page into the complete Stitch workflow for further refinement.

In the CC product of Google Labs, the Decisions API is used to quickly select appropriate emoji reactions from over 200 options. Internally in Google Search, Google also uses this API to build a security detection layer for the MCP tool, identifying potential adversarial behaviors and content.

Reference link: Gemini Decisions API Official Blog