Gemini Decisions scores bounded choices without generating prose
Google’s Gemini Decisions API scores predefined options against a shared multimodal input instead of generating an open-ended answer. One request can combine yes/no checks, multiple-choice questions and ordered rubrics, returning probabilities and confidence values. This targets routing, grading and policy checks; the application must still define the available choices and decide when uncertainty needs a fallback.
Google DeepMind · Nov 11, 2026
Why it matters
A decision endpoint can make a narrow control step explicit rather than parsing a generated sentence into an action.
More context
Try Gemini Decision 1 through a Google AI Studio API key and the documented Python, JavaScript or REST interfaces. Google lists a prototyping no-cost tier and $0.10 per million input tokens, with no output-token charge. Evaluate representative examples and ambiguous cases before setting thresholds.
Keep in mind
Released API announcement with vendor-reported calibration and latency. The 150–250 ms end-to-end figures describe evaluated conditions, not every 256K-token request or workload. Confidence scores need task-specific validation; a bounded answer is not a guarantee of correctness.
Original read