Google launches Gemini Decision 1 for high-speed multi-modal scoring

A new category of AI

Google recently launched the Gemini Decisions API, powered by a new model called Gemini Decision 1. This model is designed for a specific category of AI known as System One models, which are built to perform classification and routing rather than long-form writing. These models are intended to read a state, such as a support ticket or a web page, and return calibrated probabilities over answers defined in advance.

The name for this category stems from Daniel Kahneman's concept of fast-thinking System 1. The goal is to create systems that do not write anything because agent systems can stop paying for full paragraphs when they only need a branch, a rank, or a score. This category is relatively new, with the first major models appearing only about two months ago.

By focusing on scoring rather than generation, these models solve a common efficiency problem in AI development. Instead of paying for every word a model produces, developers can use these tools to quickly sort data. This allows for more precise control over how an AI agent decides which path to take next in a complex workflow.

Multi-modal capabilities

One of the primary distinctions of Gemini Decision 1 is its ability to process a wide variety of inputs. It can read text, images, audio, video, and PDFs simultaneously. This means a trust-and-safety team can score uploaded video clips directly without needing to transcribe the audio into text first. It also allows support teams to flag angry calls by processing the audio file in a single step.

The model is built for closed-set choices, meaning it evaluates options provided by the user. It can handle various decision types, including simple yes/no questions, multiple-choice questions with up to 255 options, and complex rubric scoring. This versatility makes it a versatile tool for any application that requires a structured output from unstructured data.

The speed of these results is a key technical highlight for developers. According to reports from Interesting Engineering, the API can return probability and confidence scores in roughly 150 to 250 milliseconds. This high-speed performance is essential for real-time applications where a delay in classification could disrupt the user experience.

The competitive landscape

Google is entering a rapidly growing space with several competitors already active. OpenAI launched its Decisions API, built on GPT-6 Luna, in early October at a price of $0.10 per million input tokens. Microsoft followed shortly after with Decision-1, which is priced lower at $0.042 per million input tokens. TypeSafe's Jev also entered the mix in September.

While Google matches OpenAI's price of $0.10, it offers a broader range of media types. OpenAI's endpoint currently only takes text and images, leaving Google with a feature advantage in the multi-modal space. However, Microsoft's lower price point remains a significant factor for teams that only need to score short text, as the 2.4x price gap can add up over millions of daily calls.

The competition is currently a feature war rather than a price war. Because Gemini Decision 1 can handle more media types than its rivals, it positions itself as a more capable tool for complex data. Developers must now decide if the extra features of Google's model justify the higher cost compared to Microsoft's cheaper alternative.

Implementation and scale

For developers, the primary advantage of Gemini Decision 1 is the ability to process high volumes of questions in a single pass. One request can carry up to 1,000 questions, and users only pay for the input tokens once. This efficiency is particularly useful for running a single document through a massive battery of checks or audits.

The model also features a 256,000-token context window, which allows it to handle large amounts of information at once. Google is making the model accessible through Google AI Studio and is rolling out integrations with popular tools like LangChain, Vercel, and Pydantic AI. These integrations help developers quickly plug the scoring capabilities into existing software stacks.

Despite these capabilities, there are still questions regarding the model's calibration. No one outside of Google has yet verified if a confidence score of 0.9 actually translates to being correct nine times out of ten on specific datasets. For this reason, experts suggest running labeled samples through all three major providers before moving a project into production.