Google launched the Gemini Decisions API on November 11, 2026, designed to provide constrained judgments rather than generate textual answers. Google says end-to-end latency is 150 to 250 milliseconds for contexts of up to 16K tokens, with little variation when each request contains 1 to 100 questions. Google charges $0.10 per million input tokens and nothing for output tokens. Each request supports up to 1,000 independent questions, but Google’s performance measurements use workloads below this limit. Google recommends testing confidence estimates with relevant labeled data and retaining human review or a second check for high-risk actions.