Google’s new AI can make 100 decisions in under a second, without writing an answer
Google introduced its Gemini Decisions API on November 11, 2026, enabling software to obtain bounded judgments from AI without generating written responses.
Powered by Gemini Decision 1, the service supports yes/no questions, choices from supplied lists, and scores against predefined rating scales. It's available through the Gemini API in Google AI Studio, marking a significant shift in how AI can be integrated directly into software workflows without the need for parsing text responses.
The practical change lies in the response format: software receives a value it can branch on, accompanied by confidence metrics, eliminating the need to parse chatbot prose. This contrasts with Google's Interactions API, which handles generated explanations, tool calls, and multi-step reasoning.
Pricing Structure
Google charges $0.10 per million input tokens with no output-token charges. Shared input is billed once per request, while each question adds its own instruction and criteria tokens. The service can process text, images, audio, video, and PDFs with up to 1,000 independent questions per request, though the total context is limited to 262,144 tokens.
Questions cannot see each other's answers, requiring code to combine results. Google reports end-to-end latency of 150-250 milliseconds for contexts up to 16K tokens, with nearly flat performance across 1-100 questions per request.
Confidence and Validation
Each choice or rating includes a confidence estimate of whether the top answer is correct. Google recommends setting thresholds using labeled examples from a developer's own traffic, weighing the costs of false positives against false negatives.
Confidence must be validated on relevant labeled data. Choice can select an option even when none fits; scope checks or an other option remain necessary. Passing a threshold does not guarantee safety; Google recommends human or second checks for high-stakes actions.
The model distributes probability among supplied options even when none fits, potentially creating specific failure cases. Google suggests implementing confidence thresholds, scope checks, or catch-all options to address this limitation.
Current Applications
Google reports that Gemini Notebook already uses the service to choose model tiers and thinking levels, achieving up to 40% lower latency for everyday queries while reserving deeper reasoning for complex requests. Search also employs the API to detect adversarial behavior and content as a safety layer for its tools.
The documentation warns that user-supplied text can argue for its own classification, recommending explicit criteria, adversarial tests, and human oversight for critical decisions such as payouts.
Related
