Last updated: 11/16/2026Last verified: 2026-11-12
Google's Gemini Decisions API, powered by the Gemini Decision 1 model, evaluates up to 1,000 yes/no, multiple-choice, and rubric-scored questions in a single pass across text, images, audio, video, and PDFs, returning calibrated probabilities in 150-250 ms for classification, routing, policy checks, and agent control.
Google''s Gemini Decisions API, powered by the Gemini Decision 1 model, evaluates up to 1,000 yes/no, multiple-choice, and rubric-scored questions in a single pass across text, images, audio, video, and PDFs, returning calibrated probabilities in 150-250 ms for classification, routing, policy checks, and agent control.
-
Answers up to 1,000 bounded yes/no, choice, and rubric-scored questions against one shared multimodal input
-
Scores candidate options in a single forward pass with calibrated probabilities and 0-1 confidence for routing ambiguous cases
-
Processes text, images, audio, video, and PDFs up to 256K tokens with tunable image and video resolution
-
Returns results in 150-250 ms end-to-end latency at $0.10 per 1M input tokens with zero output-token billing
Details
Platform
Features
- Answers up to 1,000 bounded yes/no, choice, and rubric-scored questions against one shared multimodal input
- Scores candidate options in a single forward pass with calibrated probabilities and 0-1 confidence for routing ambiguous cases
- Processes text, images, audio, video, and PDFs up to 256K tokens with tunable image and video resolution
- Returns results in 150-250 ms end-to-end latency at $0.10 per 1M input tokens with zero output-token billing
Languages
Best for
Known limitations
- Developer API for bounded decisions over closed option sets, not a general-purpose chat, writing, or generation model
- Shared input is billed once per request, but long multimodal contexts (up to 256K tokens) can drive input-token costs
- Built-in playground apps and docs live inside Google AI Studio; support inside Gemini in Google Cloud is still listed as coming soon
- Contexts supporting near-flat latency from 1 to 100 questions are documented up to 16K tokens; very large inputs may behave differently








