Google launches Gemini Decisions API for high-speed branching logic

A model built for logic, not prose

Google has introduced the Gemini Decisions API, powered by a new specialized model called Gemini Decision 1. Unlike standard large language models that generate text word by word, this model is designed to provide calibrated probabilities for specific questions. It is built for closed-set decisions where the goal is to choose between known options rather than creating new content.

The API allows developers to submit one shared input and a map of up to 1,000 questions simultaneously. It then returns a probability for each question, which the user's code can use to branch into different actions. This design targets specific tasks like routing, policy checks, tool selection, and grading within AI agents.

Because the model does not generate text, there are zero output tokens billed. This makes it a highly efficient way to handle the small, frequent logic calls that often slow down complex AI systems. As MarkTechPost reported, the model handles multimodal inputs like video and audio in a single prefill pass.

How the scoring system works

The Gemini Decisions API handles three specific types of questions to help developers structure their logic. The first is 'noul' for yes/no questions, which returns a probability bounded between 0.01 and 0.99. The second is 'choice,' which allows the model to pick from 2 to 255 different options and return normalized probabilities for each.

The third type is 'score,' which allows the model to rate an input against an ordered rubric with up to 255 levels. Every answer provided by the model includes a confidence value from 0 to 1. Google suggests that developers can use these confidence scores to identify cases that should be routed to a human reviewer or a fallback system.

Technically, the model scores these candidate options in one forward pass rather than decoding them one by one. This architecture allows the model to maintain a median compute time of 50-100 ms, with end-to-end results appearing in 150-250 ms. These speeds remain nearly flat even as the number of questions increases from 1 to 100.

Benchmarks and cost efficiency

Google tested Gemini Decision 1 against three rivals, including TypeSafe Jev 1.14 and GPT-6 Luna. The model led 7 of 9 accuracy benchmarks and showed a significant advantage in calibration. Calibration refers to how closely the model's predicted probabilities match actual outcomes, which is vital for setting automation thresholds.

The model achieved a calibration error of 0.0250, which is about half the error of Jev 1.14. In terms of pricing, the API costs $0.10 per 1M input tokens. While this is more expensive than Jev's $0.042 per 1M tokens, it is nearly 5x cheaper than the Flash-Lite model across the shared suite.

The shared input is billed only once per request, regardless of how many questions are asked. Each question only adds the cost of its own instruction and criteria tokens. This pricing structure rewards developers who bundle many logic checks into a single API call rather than making multiple separate requests.

Real-world applications

Google is already utilizing this technology internally to improve its own products. For example, Gemini Notebook uses the model to route queries to different model tiers, which helps serve everyday queries with up to 40% lower latency. This demonstrates how the model can optimize the backend of a large-scale AI service.

Other practical uses include Google Labs' CC, which uses the model to pick emoji reactions from a list of more than 200 options. Additionally, Stitch's rapid canvas uses the API to build playable wireframes in seconds. These examples highlight how the model excels at picking from a predefined set of options based on complex inputs.

The API is currently available through the Gemini Developer API v1beta and requires the google-genai SDK 3.2 or later. While it is not currently available for self-hosting or as an open-weights model, it is being integrated into major ecosystems like LangChain, Vercel AI SDK, and Pydantic AI. A no-cost tier is also available for developers who want to prototype their ideas in Google AI Studio.