Google launches Gemini Decision 1 for faster AI judgments

A shift from generation to scoring

Google has officially released Gemini Decision 1, its first model specifically engineered for making choices rather than writing long-form text. While most popular AI models work by predicting the next word in a sequence, this new model uses a single forward propagation pass to evaluate options. This means it can provide a result much faster because it does not have to decode a response word by word.

The model is designed to handle tasks within a closed set of possibilities. Developers can provide the input content, the specific questions to be judged, and the candidate options. The system then directly returns the judgment, a probability score, and a confidence level. This architecture eliminates the need for developers to extract results from a long string of generated text, which was the standard workflow for previous models.

By focusing on scoring, Google aims to create a more efficient underlying component for AI systems. The model supports multi-modal inputs, meaning it can process text, images, audio, video, and PDF files. It can handle up to 1,000 independent questions in a single request, making it suitable for high-volume classification tasks that require rapid responses.

Three core decision-making types

Gemini Decision 1 offers three distinct capabilities for developers. The first is binary judgment, or yes/no tasks. For instance, a user could upload an image to determine if it contains a safety hazard, and the model would provide the probability that the condition is true. This is useful for rule-based content moderation or identifying specific objects in a scene.

The second capability is multiple-choice selection, supporting up to 255 candidate options. Instead of just picking one answer, the model returns a normalized probability for every option provided. This allows for more nuanced data collection where the system can see how closely a piece of content aligns with various pre-set categories or labels.

The third type is ordered rubric scoring, which allows for customized grading standards. Developers can set different levels of quality or performance and the model will return continuous weighted scores along with the probability of each level. This allows for complex evaluations, such as grading a customer service call based on a specific set of instructions, in a single request.

Benchmarks and performance gains

In a series of tests reported by 36Kr, Gemini Decision 1 took the first place in 7 out of 9 benchmark evaluations. It achieved a maximum accuracy rate of 92.2% in the JevBench test, outperforming competitors like OpenAI's GPT-6 Luna and TypeSafe Jev 1.14. It also showed significant leads in medical Q&A tests and multilingual evaluations, demonstrating its ability to handle diverse data types accurately.

One of the most important metrics for developers is probability calibration, which measures how well the model's confidence scores match its actual accuracy. Gemini Decision 1 showed a calibration error of only 0.0250 in the Top-1 ECE test. This is about 65% lower than GPT-6 Luna and 88% lower than Gemini 3.5 Flash-Lite, which is critical for systems that need to decide when to hand off a task to a human.

Speed and cost are also central to the model's appeal. The model maintains a stable latency of 150 to 250 milliseconds even when the number of questions in a request increases from 1 to 100. Furthermore, the cost of the Gemini Decisions API is nearly 5 times lower than calling Gemini 3.5 Flash-Lite via the Interactions API, making it a highly aggressive pricing option for high-volume use cases.

Real-world internal applications

Google is already integrating the Gemini Decisions API into its own products to improve performance. In Gemini Notebook, the API acts as a real-time model router. It analyzes the complexity of a user's request to decide which model tier to use, allowing simple questions to be answered by cheaper, faster models while routing complex tasks to more capable systems. This dynamic routing has reportedly reduced the latency of daily queries by up to 40%.

Other internal applications include Google Stitch, where the API helps generate interactive product flow wireframes in seconds based on voice or click instructions. In Google Labs, the model is used to select appropriate emoji reactions from a list of over 200 options. These examples show how the model can handle both creative design tasks and simple interactive elements within a single framework.

Finally, Google is using the technology to build a safety detection layer within Google Search. This layer identifies potential adversarial behaviors and content for MCP tools. By using a dedicated decision-making model for these safety checks, Google can maintain a high level of security without the overhead of a full generative model for every single interaction.