Google Releases Gemini Decision 1: Skips Token-by-Token Generation, Evaluates 1,000 Questions in a Single Request

Author: Dongmei, AI Frontline

While the entire AI industry is still competing on who can generate more tokens or has stronger reasoning capabilities, Google has introduced a new model that does not rely on word-by-word answer generation.

Just now, Google officially launched Gemini Decision 1 and simultaneously introduced 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, it scores predefined candidate answers in a single forward pass.

Google CEO Sundar Pichai also posted on the X platform introducing 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 brand-new Gemini Decision 1 model.

Unlike traditional models that generate text token by token, this model is specifically designed for decision-making tasks within a closed set of options. It can score up to 1,000 questions across text, images, audio, video, and PDFs in a single processing step.

Here are several noteworthy details:

Supports three types of decisions: yes/no judgment, multiple choice (up to 255 options), and scoring based on grading rubrics;

Priced at only $0.10 per million input tokens, with zero cost for output tokens, making its cost roughly one-fifth of Flash-Lite;

Takes the lead in multiple benchmark evaluations while maintaining the lowest calibration error.

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

He also demonstrated Stitch Rapid designing a complete product workflow in seconds, played at normal 1x speed.

This introduction reveals the biggest difference between Gemini Decision 1 and traditional large models: it is not responsible for articulating the answer, but rather for making judgments from existing options.

In the past, when developers wanted AI to determine whether an email was spam or decide which agent a task should be assigned to, they typically had to invoke a generative large model and then extract results from the generated text. Gemini Decision 1 eliminates the word-by-word decoding step.

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

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

According to Google's published data, Gemini Decision 1 can handle 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 to 250 milliseconds.

Pricing is also quite aggressive: only $0.10 per million input tokens, with no charge for output tokens. In the same group of benchmark tests 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 applied it to scenarios such as AI model routing, agent tool selection, product design, and security detection, attempting to make it the underlying component in AI systems specifically responsible for judgment and decision-making.

Processing 1,000 Questions in One Go, with Up to 255 Options

According to Google's released technical documentation, Gemini Decision 1 offers 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 if a piece of text violates established rules. The model provides the probability of each condition being met.

The second type is Multiple Choice. Developers can preset 2 to 255 candidate options. The model not only selects the best-matching answer but also returns normalized probabilities for each candidate.

The third type is Ordered Rubric. Developers can define grading standards for different levels themselves, allowing the model to evaluate the input content and return continuous weighted scores along with probabilities for each level.

These three types of tasks can also be executed concurrently in a single request.

Google provided a simple example: uploading a photo of an office breakroom countertop to the model, accompanied by a text description, requesting it to simultaneously judge whether there is a liquid spill on the surface, 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 judgments.

Each decision provides a confidence score between 0 and 1, facilitating the system to hand uncertain results over to other models or for manual review.

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

According to official Google descriptions, Gemini Decision 1 leads in benchmark evaluations covering agent routing, reasoning, ranking, and multimodal understanding, while achieving the lowest calibration error.

In test contexts not exceeding 16K tokens, its median compute-phase latency remains between 50 and 100 milliseconds, with an end-to-end latency of approximately 150 to 250 milliseconds. Even when the number of questions in a single request increases from 1 to 100, the latency remains relatively stable.

On the shared benchmark dataset used officially, the cost of the Gemini Decisions API is nearly 5 times lower than invoking Gemini 3.5 Flash-Lite via the Interactions API.

Ranking First in 7 Out of 9 Tests, with Accuracy Peaking at 92.2%

Alongside this launch of Gemini Decision 1, Google also released a highly noteworthy model comparison report card.

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

Looking at the results, Gemini Decision 1 took first place in 7 out of 9 benchmark tests and also led in probability calibration, but in two other tests and overall invocation cost, competitors still performed 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 scored 58.7% Macro-F1, 4 percentage points higher than GPT-6 Luna and 11.5 percentage points higher than Flash-Lite.

The gap widened further in the New Yorker Caption test. Gemini Decision 1 scored 79.7%, compared to GPT-6 Luna's 68.4%, a difference of 11.3 percentage points. This test involves understanding and selecting comic captions, testing the model's contextual comprehension and candidate evaluation capabilities.

In the medical QA test PubMedQA, Gemini Decision 1 also ranked first with a score of 79.8%, higher than GPT-6 Luna's 74.4%.

Results for 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 an outright lead across the board.

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

In the When2Call test evaluating tool invocation judgment capabilities, TypeSafe Jev 1.14 held a clearer advantage with a score of 82.0%, compared to Gemini Decision 1's 79.9%.

This result is particularly noteworthy. Tool selection and timing judgment for invocation are key agent application scenarios for Gemini Decision 1, yet in this test, the dedicated decision model TypeSafe Jev maintained its lead.

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

More noteworthy than pure accuracy is the probability calibration capability that Google specifically emphasized this time.

In the Top-1 ECE (Expected Calibration Error) test, Gemini Decision 1's calibration error was only 0.0250, compared to 0.0721 for GPT-6 Luna.

In comparable comparisons, TypeSafe Jev 1.14 scored 0.0494, while Flash-Lite reached 0.2104. In other words, Gemini Decision 1's calibration error is approximately 65% lower than GPT-6 Luna's and roughly 88% lower than Flash-Lite's.

This metric is particularly important for agent systems.

For example, when an agent needs to determine whether a piece of code contains a security vulnerability, the model provides not only a "present" or "absent" result but may also include a probability value. The system can decide whether to automatically take action or hand it over to human reviewers based on the relationship between the probability and actual accuracy rates.

Google Already Using It: Agent Routing Latency Reduced by Up to 40%

Unlike some new models that only showcase benchmark results, Google also disclosed use cases of the Gemini Decisions API in its internal products this time.

The most notable among them is Gemini Notebook.

Google is leveraging the Decisions API as a real-time model router, selecting the appropriate model tier and reasoning intensity based on the complexity of user requests. For daily simple questions, the system prioritizes calling lower-cost, faster-responding models; for complex reasoning tasks, it routes the request to more capable models.

Google states that this dynamic routing approach can reduce latency for routine queries by up to 40%.

This design closely aligns with current trends in agent systems. When an agent completes a task, it often needs to continuously judge the next steps, such as selecting which tool to use, invoking which model, or deciding whether human intervention is needed.

If every simple decision had to be handed to large generative models, the system would incur additional latency and costs. Dedicated decision models offer an alternative: entrusting relatively straightforward, clearly optioned judgments to low-latency models, while reserving complex tasks for general-purpose large models.

Beyond Gemini Notebook, Google also revealed three other application directions.

In Google Stitch, the newly launched Rapid Canvas experimental feature utilizes the Decisions API to generate interactive product workflow wireframes in seconds, rapidly responding to user modification commands via voice or clicks. After users confirm the design direction, they can transition the page to the full Stitch workflow for further refinement.

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

Reference link: Official Blog Post on Gemini Decisions API