Google DeepMind has released the Gemini Decisions API, powered by a specialized model called Gemini Decision 1, designed to eliminate sequential text token generation in favor of evaluating closed-choice options in a single pass. According to Google, the system returns calibrated probabilities and confidence scores to resolve latency bottlenecks and token cost overheads in generative agent workflows. Rather than producing tokens one by one for binary decisions or routing pathways, the interface targets deterministic classification directly. The service accepts multimodal contexts of up to 256,000 tokens encompassing text, audio, images, video, and PDFs, allowing users to process up to 1,000 independent queries simultaneously against the same input in one call.

The pricing structure sets input costs at $0.10 per million tokens and eliminates output token fees entirely. Because multiple questions share the same contextual data, the underlying input payload is billed only once per request regardless of query volume. In evaluation benchmarks on the public JevBench suite, Gemini Decision 1 reduced operational costs to roughly one-fifth of the expense associated with running the Gemini 3.5 Flash-Lite Interactions API, while achieving 92.2% classification accuracy on the public split. Google reported that internal deployments such as Gemini Notebook cut daily query routing latency by up to 40% by bypassing generative delays when determining whether to trigger external tools, complex reasoning engines, or simple web searches.

The rollout intensifies competition among major cloud and AI providers seeking to capture non-generative classification infrastructure. As agentic frameworks expand, utilizing full-scale generative models for lightweight tasks such as policy compliance, safety filtering, and model routing has created severe latency and computational overhead. Competitors including Jev, OpenAI, and Cloudflare have each introduced non-generative endpoints to address this challenge. Google is seeking an advantage in this high-throughput branching market by pairing rapid classification with deep multimodal context, enabling single-pass evaluation across rich video, audio, and document feeds.