Google Launches Gemini Decisions API for Low-Latency Multimodal Classification

Google has unveiled the Gemini Decisions API, a purpose-built endpoint for fast, closed-set classification tasks that sits alongside the generative Gemini API in AI Studio. The underlying Gemini Decision 1 model evaluates a single multimodal input — up to 256K tokens spanning text, images, audio, video, and PDFs — against as many as 1,000 independent questions in one forward pass, returning calibrated confidence scores rather than generated tokens.

Three decision primitives are supported: yes/no probability estimates, multiple-choice selection from 2–255 options, and rubric-based scoring on a 1–5 or 1–10 scale. Because the model scores candidates directly without token-by-token decoding, median model compute latency stays in the 50–100 ms range, with end-to-end latency of 150–250 ms that remains nearly flat whether you ask 1 or 100 questions. Pricing is set at $0.10 per million input tokens with zero output tokens billed, making it roughly five times cheaper than using Gemini 3.5 Flash-Lite with the Interactions API for equivalent routing, grading, or policy-check workloads.

Why it matters for GPU / AI infrastructure

The Decisions API shifts a class of high-volume inference — routing, content moderation, agent tool selection, document triage — from generative decoding to a single-pass scoring paradigm. For infrastructure teams, this means dramatically lower compute per decision, predictable latency profiles, and the ability to batch thousands of evaluations per request without scaling token generation. Workloads that previously required fleets of GPU instances for autoregressive decoding can now be consolidated onto fewer accelerators, reducing both CapEx and OpEx for AI serving stacks.

Developers can access the API immediately via Google AI Studio with Python, JavaScript, and REST client libraries. The native multimodal architecture also allows mixing modalities — such as pairing a call recording with account metadata or a scanned claim form with damage photos — while tuning image and video resolution to control token cost.