Gemini Decision 1 and Decisions API Release Notes
Overview
Google has introduced the Gemini Decisions API, powered by the new Gemini Decision 1 model. This service is designed for high-throughput, low-latency classification tasks, allowing developers to evaluate up to 1,000 questions—including yes/no, multiple-choice, and rubric-based scoring—against multimodal inputs in a single forward pass. By bypassing token-by-token generation, the API achieves end-to-end latency of 150-250 ms.
Core Capabilities and Architecture
The Gemini Decision 1 model is built on a native multimodal architecture, enabling it to process text, images, audio, video, and PDF documents simultaneously. Unlike standard generative models that output text, the Decisions API returns probabilities and confidence scores for defined categories.
Key decision types include:
- noul (Yes/No): Estimates the probability of a condition being true.
- choice (Multiple Choice): Selects from 2 to 255 options with normalized probabilities.
- score (Ordered Rubric): Assigns a continuous weighted score based on a predefined rubric.
Because the model does not perform token-by-token decoding, it is significantly more efficient for classification tasks. Users are billed $0.10 per 1M input tokens, with no charges for output tokens, making it a cost-effective alternative for high-volume decision-making workflows.
Industry Applications
Google is currently utilizing the Decisions API across several internal products to serve as a low-latency control plane:
- Model Routing: Gemini Notebook uses the API to classify user intent and route queries to the appropriate model tier, reducing latency for simple tasks.
- Design Iteration: The Stitch design tool uses the API to process voice and tap inputs to update wireframes in real-time.
- Safety and Moderation: Google Search employs the API to detect adversarial content and behavior within strict latency constraints.
Community Insights and Discussion
Developer reception has focused on the API's unique multimodal support and its competitive performance against existing decision-making models. While some users noted that the pricing is higher than specialized alternatives, the ability to ingest audio and video directly is seen as a significant differentiator.
"This appears the first decision AI that actually handles audio, video, and PDF, and I'm surprised this isn't highlighted everywhere... I doubt anyone's picking it for the +2% higher points, but being able to answer based on a PDF or call is groundbreaking."
Some community members raised questions regarding the necessity of a dedicated decision API versus using standard LLMs, while others expressed concerns about the lack of acknowledgment regarding the influence of earlier "Jev"-style decision APIs. Technical users also noted the importance of the confidence scores provided by the API, which allow for automated routing to human reviewers when the model's confidence is low.
Sources
Related
- Dispatch
- Dispatch
- Dispatch
- Dispatch
- Dispatch