A Look at Gemini Decision 1
Google has announced the launch of the Gemini Decisions API, a new API based on the Gemini Decision 1 model, aimed at enabling developers to build fast, low-cost decision-making systems. This model is uniquely capable of evaluating up to 1,000 independent questions about a single piece of content in a single API call, making it ideal for tasks that require multiple analyses on the same data.
What Sets This Model Apart Technically?
Unlike traditional generative AI models that generate answers word-by-word, Gemini Decision 1 evaluates available options and calculates their probabilities all at once. The model supports the processing of text, images, audio, video, and PDF files, with a context window of up to 256,000 tokens. It also provides developers with a "confidence score" ranging from 0 to 1 for each answer, allowing results with low confidence to be routed for human review or to more complex models.
In terms of performance, Google indicates that the model is capable of processing data up to 16,000 tokens in a time ranging from 150 to 250 milliseconds, with a computational time ranging from 50 to 100 milliseconds.
Why Does This News Matter?
This launch is a strategic step toward reducing operating costs in applications that require frequent checks on the same documents. The pricing structure is based on $0.10 per million input tokens, with no charges for output tokens. Most importantly, the cost of the shared content is calculated only once per call, which significantly reduces costs compared to traditional models like Gemini 3.5 Flash-Lite, with Google claiming the cost is approximately 5 times lower.
Use Cases and Availability
Google has already begun integrating this technology into its products, such as Gemini Notebook to determine the level of reasoning required for queries, and the Stitch tool for prototyping, in addition to a security layer in the Google search engine to detect malicious behaviors. The model is currently available via Google AI Studio, with integration support for platforms such as LangChain, Vercel, and Pydantic AI, and is planned to be available soon on Google Cloud.