Google gives developers a purpose-built API for bounded decisions

Google has introduced the Gemini Decisions API, powered by Gemini Decision 1, for bounded choices such as yes-or-no checks, multiple-choice selection, and rubric scoring. Google says a request can share one input across text, images, audio, video, and PDFs, then ask as many as 1,000 independent questions about it. The model returns probabilities and confidence scores rather than composing a conventional answer token by token. That makes the offering aimed at classification and routing steps inside an application, where a compact decision is useful and free-form explanation is not the main output. The API is available through the Gemini API in Google AI Studio, according to Google’s announcement.

The design changes the unit of work from one generated answer to many scores against a common input. Google describes use cases including tool selection, policy checks, model routing, and context grading, and says inputs can reach 256K tokens. Its examples combine unlike materials, such as a call recording with account metadata or a scanned claim with damage photos. For developers already using a generative model to label or triage material, batching these questions into one call could simplify orchestration. The announcement does not establish that the model can replace a full agent or explain complex decisions; the endpoint is deliberately for closed-set choices.

Google reports 150 to 250 milliseconds end-to-end latency and 50 to 100 milliseconds of model compute, with latency staying nearly flat from one to 100 questions in contexts up to 16K tokens. It also says the API costs $0.10 per million input tokens and bills no output tokens, and claims it was nearly five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API across a shared benchmark suite. These are Google’s published measurements and comparison. The company says its model led its tested suite across agent routing, reasoning, ranking, and multimodal understanding with the lowest calibration error, but the announcement’s results do not show how performance transfers to a particular team’s data, label balance, or failure costs.

The practical appeal is clear when one expensive or slow generative step currently answers many small questions about the same document or media item. Confidence scores can help an application send uncertain cases to a larger model or a person, while image and video resolution controls give developers a way to trade visual detail against token cost. Still, confidence is not a guarantee of correctness, and policy checks or routing decisions need evaluation against representative edge cases before they control consequential actions. Google’s announcement gives headline timing, pricing, and benchmark comparisons, but no independent results or workload-specific reliability guarantees. Teams should measure end-to-end behavior, including fallback rates and the cost of mistakes, before treating the advertised savings as their own.