Google launches Gemini Decision 1 for high-speed scoring

A shift toward scoring

Google has introduced a new model called Gemini Decision 1, which is specifically designed for bounded choices. Unlike standard large language models that generate text one word at a time, this model scores candidate options in a single forward pass. This means it can evaluate a wide range of possibilities almost instantly.

The model is intended for tasks like tool selection, policy checks, model routing, and content grading. Because it focuses on scoring rather than generation, it provides probabilities and a confidence score between 0 and 1 for every answer. This allows developers to build systems that know when they are uncertain and can hand off those cases to a human or a different model.

As MLQ.ai reported, Google is making this available through the Gemini API in Google AI Studio. The model is already being used internally by Google to power various features, including a safety layer for search tools and a design experiment that turns voice edits into wireframes.

How the scoring works

The technical architecture of Gemini Decision 1 allows it to process up to 1,000 independent questions in one request. It handles multimodal inputs, meaning it can look at text, images, audio, video, and PDFs simultaneously. A single request can contain up to 256,000 tokens of this shared input data.

Developers can mix three specific types of questions in a single call. These include yes/no checks to see if a condition holds, multiple-choice questions with up to 255 options, and ordered rubric scores. The rubric scores are particularly useful because they return a weighted score along with probabilities for each level of the rubric.

Speed is a primary feature of this new model. Google says the end-to-end process takes between 150 and 250 milliseconds. The model's median compute time is even faster, ranging from 50 to 100 milliseconds on contexts up to 16,000 tokens, and it remains relatively flat even as the number of questions increases.

Cost and performance

Pricing for Gemini Decision 1 is set at $0.10 per million input tokens, with no charge for output tokens. Because the shared input is billed only once per request, the cost remains the same regardless of how many questions are asked. This structure makes it nearly five times cheaper than Gemini 3.5 Flash-Lite on Google's shared benchmark suite.

Google claims the model leads its benchmark suite, which covers areas like reasoning, ranking, and multimodal understanding. One of the key metrics highlighted is the model's low calibration error. This refers to how accurately the model's predicted probabilities match the actual likelihood of an outcome.

The model is designed for high-volume work where efficiency is critical. By avoiding the token-by-token generation of standard models, it reduces both the time and the money required to perform complex evaluations. This makes it a competitive option for companies building large-scale AI agents.

Real-world applications

Google is already integrating the API into several products to improve user experience. For example, Gemini Notebook uses the API to route queries to different model tiers, which Google says reduces latency by up to 40% on everyday queries. Another tool, Stitch, uses it to create wireframes almost instantly from voice or tap edits.

In the realm of content and safety, the model is being used to detect adversarial behavior in search tools. It can also be used for simple tasks like picking an emoji reaction from a list of more than 200 options. These examples show how the model can handle both high-level reasoning and simple, high-speed classification.

For developers, the model offers a way to handle complex scenarios in a single step. Google provided an example where a photo of a breakroom counter is checked for a spill, matched to a cleanup action, and rated for severity all in one request. This multi-step logic is executed much faster than if each step were a separate request to a generative model.

The broader landscape

Google is not the only company moving toward this type of scoring model. Other major players like OpenAI, Cloudflare, and Microsoft have released similar offerings in recent weeks. These models follow a trend started by the startup Jev, which introduced a version of scoring technology first.

The availability of these models is expanding quickly across the AI ecosystem. Integrations with popular platforms like OpenRouter, LangChain, and Vercel are currently rolling out. This suggests a growing demand for specialized models that prioritize judgment and classification over long-form creative writing.

The pricing of Gemini Decision 1 also aligns with current market trends. It matches the $0.10 per million input token price set by Meta for its Muse Ember model. While Muse Ember is a text generator, the similar pricing suggests a standard for high-volume, specialized AI tasks is beginning to form.