Google LLC today launched Gemini Decisions API, an artificial intelligence service designed to answer user questions in a cost-efficient manner.
It’s powered by a new AI model called Gemini Decision 1.
Traditional large language models are designed to answer a broad range of prompts. Gemini Decision 1 has a much narrower focus: It can process just three types of questions. The upside is that it answers those questions at a significantly lower cost. According to Google, the model is nearly five times cheaper than its entry-level Gemini 3.5 Flash-Lite LLM.
The model can answer yes-no questions, multiple choice questions and prompts that ask it to rate an item on a user-specified scale. A cybersecurity team, for example, could ask the model to rate the severity of an antivirus alert on a scale of 1 to 10.
Gemini Decision 1 prompts include two items: the user’s questions and the dataset to which those questions pertain. The dataset can include up to 256,000 tokens worth of text, audio, images, video or some combination thereof.
LLMs usually generate verbose free-flowing text in response to user questions. Gemini Decision 1, in contrast, outputs terse answers such as “yes” or “no.” It also displays a score that indicates how confident it is in the accuracy of a given response. That score helps applications decide whether they should make use of the output.
Neural networks designed to generate brief answers to simple questions are known as decision models. The first such model, Jev, was launched in mid-September by a startup called TypeSafe AI Inc. It gained a million users within a few days. Three weeks after the launch, TypeSafe raised $870 million at a $7.5 billion valuation.
Decision models can answer the narrow set of questions for which they’re optimized faster than LLMs. Additionally, they’re easier to integrate into applications. An application that powers some of its features using an AI model must check the model’s output for errors and filter redundant details before ingesting it. Decision models’ concise, structured responses make both tasks simpler.
“Gemini Decision 1 scores options without token-by-token decoding, keeping median latency 50-100 ms of model compute (150-250 ms end-to-end), across multimodal inputs and contexts up to 16K tokens, remaining nearly flat from 1 to 100 questions per request,” Google DeepMind product manager Min Choi and principal scientist Felix Yu wrote in a blog post.
Google checked how Gemini Decision 1 compares to Jev and OpenAI Group PBC’s Decisions API, a competing service that launched last month. The search giant says that its model performed better across seven of the nine benchmarks it tested. However, Jev proved to be slightly more cost-efficient.
Gemini Decision 1 is available through the Gemini Decisions API for 10 cents per 1 million input tokens. Google doesn’t charge for output tokens. The algorithm is also accessible through a limited free tier in Google AI Studio, a service that developers can use to familiarize themselves with the search giant’s models.
Image: Google