Google DeepMind announced the public availability of the Gemini Decisions API, an inference endpoint engineered specifically for high-speed, bounded classification and decision-making. Unveiled by Product Manager Min Choi and Principal Scientist Felix Yu, the system is driven by a specialized model designated Gemini Decision 1 (gemini-decision-1) and accessible immediately through Google AI Studio.
Unlike conventional large language models that generate responses token by token via autoregressive decoding, Gemini Decision 1 computes normalized logit distributions across pre-defined candidate options in a single forward pass. This architecture allows client applications to submit complex multimodal context—encompassing text, audio recordings, high-resolution imagery, video clips, and PDF documents up to 256,000 tokens—paired with up to 1,000 discrete questions in one API call.
The API exposes three core decision primitives: 'noul' for calibrated Boolean verifications with optional true/false criteria; 'choice' for selecting among 2 to 255 discrete options with exhaustive normalized probability distributions; and 'score' for grading against an ordered rubric, returning continuous weighted scores. Every decision returns a calibrated confidence score between 0 and 1, enabling agentic orchestrators to establish deterministic threshold gates and route low-confidence edge cases to secondary reasoning tiers or human oversight.
Performance evaluations published by Google demonstrate median model compute latencies between 50 and 100 milliseconds, translating to 150 to 250 milliseconds end-to-end execution. Because candidate scoring is calculated concurrently, response latency remains virtually flat whether processing a single question or scaling to 100 concurrent queries over a shared multimodal payload.
The economic model represents a substantial shift for agent architectures. Google is billing Gemini Decision 1 at $0.10 per million input tokens with zero output token charges, zero transcription surcharges, and no variable thinking-token fees. Across standardized benchmark suites evaluating agent routing, safety filtering, and retrieval ranking, Google calculates that the endpoint delivers approximately fivefold cost reductions compared to invoking Gemini 3.5 Flash-Lite through standard generative APIs.
Internal product teams have already integrated the endpoint as an operational control plane. Gemini Notebook leverages Decisions API to route queries dynamically to optimal model tiers, reducing routine prompt latency by up to 40%. The Stitch UI design platform uses it to synthesize playable wireframes in real time, while Google Search employs the engine to evaluate Model Context Protocol (MCP) tool execution safety against prompt injection datasets.