Gemini Decisions API: Fast, calibrated multimodal decisions
Gemini Decisions API scores up to 1,000 bounded questions on shared multimodal input, giving coding agents a low-latency option for routing, policy checks, and context grading.
Google’s **Gemini Decision 1** scores up to **1,000 questions** in one pass over shared text, image, audio, video, or PDF input. The API returns probabilities and confidence scores for yes/no, multiple-choice, and rubric questions; Google reports **150–250 ms** end-to-end latency and pricing of $0.10 per 1M input tokens with no output-token charge.
For coding agents, consider it for bounded decisions such as tool selection, model routing, policy checks, or grading context. A confidence score can help route uncertain cases to a fallback or human review, while one shared input can support several checks.
Google’s **Gemini Decision 1** scores up to **1,000 questions** in one pass over shared text, image, audio, video, or PDF input. The API returns probabilities and confidence scores for yes/no, multiple-choice, and rubric questions; Google reports **150–250 ms** end-to-end latency and pricing of $0.10 per 1M input tokens with no output-token charge. For coding agents, consider it for bounded decisions such as tool selection, model routing, policy checks, or grading context. A confidence score can help route uncertain cases to a fallback or human review, while one shared input can support several checks. The latency and cost comparisons are Google’s claims; the material does not include independent results or details about how these figures vary in production. The API is available through Google AI Studio, with broader integrations rolling out.
This adds a direct Google API option for bounded agent decisions, with batch questions over shared multimodal input and returned probabilities and confidence scores. It makes confidence-based fallback a concrete integration possibility alongside specialized routers, while the latency and price figures remain vendor claims rather than independent evidence.