Google launches Gemini Decisions API, dedicated to judgment rather than text generation
Google today launched Decisions API in Gemini API, together with the new model Gemini Decision 1. The interface does not generate text answers, but instead scores the input and a set of questions, directly returning the answer to each question and its confidence.
A single request can carry one multimodal input (text, images, audio, video, PDF, up to 256K tokens) and up to 1,000 separate questions. The model scores candidate options in a single forward pass, returning probabilities and confidence, with end-to-end latency of about 150-250 milliseconds. Pricing is US$0.1 per 1 million input tokens, and output tokens are free.
Applicable scenarios include tool selection, policy checks, model routing and context scoring. Officially, it is said to lead in benchmarks for agent routing, reasoning, ranking and multimodal understanding, and to have the lowest calibration error. Compared with full text generation, this path is aimed at closed-set judgment and avoids token-by-token decoding.
Original source: blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api
ABAB AI Insight
The model provider is separating closed-set judgment from the text generation path and introducing a dedicated decision interface. Google has launched the Gemini Decisions API and the Gemini Decision 1 model, which support scoring up to 1,000 questions in a single request and returning probabilities and confidence rather than generating text. The input covers multimodal content, up to 256K tokens, and output tokens are free.
Unlike generating answers token by token, this path completes option scoring in a single forward pass, with latency kept at 150-250 milliseconds. In similar cases, capital is more focused on how judgment speed and cost affect branch calls in agent workflows, and whether confidence can support fallback for low-confidence cases. Official data show that it has the lowest calibration error across multiple benchmarks, and the cost is about one-fifth of that of the full interaction API.
This launch extends Gemini from general-purpose generation to a dedicated decision primitive. Beneficiaries are agent and application developers who need high-frequency, low-latency branch judgments; full generation models will continue to handle open-ended response tasks. The structural change is that closed-set judgment has been separated from text generation, forming a low-cost, zero-output-fee dedicated path that supports scaled calls for tool selection and policy checks.
Cognitive Laws
A dedicated decision model scores options in a single pass, avoiding the latency and output fees of text generation.
A single request supports up to 1,000 questions, turning branch judgments from sequential calls into batch scoring.
Zero output fees and low latency make closed-set judgment a low-cost primitive for agent workflows.