Overview
Gemini Decision 1 is a multimodal model for fast, closed-set decisions through the Gemini Decisions API. Instead of generating answer text token by token, it scores candidate outcomes in a single forward pass and returns probabilities and confidence scores. Google announced it on November 11, 2026. (blog.google)
Capabilities
- Accepts shared inputs combining text, images, audio, video, and PDFs, up to 256K tokens, with up to 1,000 independent questions per request.
- Supports three question types:
noulfor yes/no probabilities;choicefor selecting among 2–255 options with normalized probabilities; andscorefor an ordered rubric with a continuous weighted score and per-level probabilities. - Each answer includes a confidence score from 0 to 1. Image and video resolution can be adjusted to balance visual detail and token cost. (blog.google)
Uses and performance
Google identifies tool selection, policy checks, model routing, and context grading as intended uses. Low-confidence results can be routed to a fallback or human reviewer. Google reports median model-compute latency of 50–100 ms and end-to-end latency of 150–250 ms for multimodal inputs and contexts up to 16K tokens; these are Google’s reported measurements, not guarantees for all workloads. (blog.google)
Limitations
The model is designed for bounded decisions rather than open-ended text generation. A request is limited to 1,000 independent questions and a shared input of up to 256K tokens; multiple-choice questions allow 2–255 options. Confidence scores should not be treated as guarantees of correctness. At announcement, Google Cloud support was described as coming soon. (blog.google)
Pricing and availability
Google’s announced API price is USD $0.10 per 1 million input tokens, with no output-token charges. Shared input is billed once per request; each question adds its own instruction and criteria tokens. Google says there are no separate thinking-token, audio transcription, or OCR charges, and offers a no-cost API tier for prototyping in Google AI Studio. (blog.google)