Google Launches Gemini Decision 1 to Compete in Specialized AI Model Market

Google LLC unveiled Gemini Decisions API, a fresh artificial intelligence offering engineered to handle user inquiries while keeping expenses down. The service runs on Gemini Decision 1, a purpose-built AI model that takes a fundamentally different approach from conventional large language models.

While traditional LLMs tackle virtually any type of question, Gemini Decision 1 operates within strict boundaries. The model handles only three categories of queries: yes-no questions, multiple choice selections, and rating tasks on user-defined scales. This specialization delivers a major advantage—the model costs roughly one-fifth the price of Google's entry-level Gemini 3.5 Flash-Lite offering.

Consider a security operations team evaluating an antivirus alert. They could feed Gemini Decision 1 the alert details and ask it to score the threat severity from 1 to 10. Each prompt consists of two components: the question itself and the relevant dataset. That dataset can accommodate as much as 256,000 tokens of text, audio, images, video or any combination.

Where standard LLMs produce lengthy, flowing prose, Gemini Decision 1 delivers compact responses—a simple "yes," "no," or numerical rating. The model also supplies a confidence score reflecting how certain it is about the answer, allowing applications to determine whether to trust the output.

Neural networks engineered to provide brief answers to focused questions fall into a category called decision models. TypeSafe AI Inc., a startup, released the first such model, Jev, in mid-September. The product attracted one million users within days, and three weeks later TypeSafe closed a funding round at a $7.5 billion valuation on $870 million raised.

Decision models offer distinct advantages over conventional LLMs. They respond faster to the specific questions they're built for, and their structured, concise outputs integrate more easily into applications. Since decision models produce tidy answers rather than verbose text, developers spend less effort validating results and removing extraneous information before using them.

Google DeepMind product manager Min Choi and principal scientist Felix Yu described the model's performance characteristics in a blog post: "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 benchmarked Gemini Decision 1 against two rivals: Jev and OpenAI Group PBC's Decisions API, which arrived on the market last month. Across nine test scenarios, Google's model outperformed the competition on seven. Jev, however, maintained a slight edge on cost efficiency.

Gemini Decision 1 is accessible via the Gemini Decisions API at a rate of 10 cents per 1 million input tokens, with no charges for output tokens. Developers can also explore the model at no cost through a restricted free tier within Google AI Studio, Google's platform for experimenting with its AI offerings.