谷歌推出 Gemini Decisions API,该服务基于新模型 Gemini Decision 1,旨在为文本、图像、音频、视频及 PDF 等多模态输入提供快速且校准的有界选择能力 1。在单次前向传递中,该 API 即可完成工具选择或政策检查等任务,端到端延迟低至 150-250 毫秒,模型计算的中位延迟仅为 50-100 毫秒 1。此外,其定价策略极具竞争力,输入 token 价格为每百万 $0.10,而输出 token 完全免费,整体成本比使用 Interactions API 的 Gemini 3.5 Flash-Lite 便宜近 5 倍 1。
Gemini Decision 1 支持三种核心决策类型:noul(是/否判断)、choice(多选)以及 score(有序评分),并已与 OpenRouter、LangChain、Vercel、LiteLLM 和 Pydantic AI 等主流生态集成 1。目前,该 API 已应用于 Google 内部产品线,例如 Gemini Notebook 利用其实时路由功能将延迟降低了 40%,Search 则将其用于 MCP 工具的安全层检测,Stitch 也借此实现快速设计迭代 1。
Google has introduced the Gemini Decisions API, a new service powered by the core model Gemini Decision 1 (gemini-decision-1) designed to deliver fast and calibrated multimodal decision-making capabilities 1. The API supports diverse input types, including text, images, audio, video, and PDFs, enabling it to handle bounded choices such as tool selection and policy checks within a single forward pass 1.
The service offers significant performance advantages, with end-to-end latency ranging from 150 to 250 milliseconds and median model computation latency between 50 and 100 milliseconds 1. From a cost perspective, the API charges $0.10 per one million input tokens while imposing no fees for output tokens, making it nearly five times cheaper than using Gemini 3.5 Flash-Lite via the Interactions API 1. It accommodates three specific decision types: "noul" for yes/no questions, "choice" for multiple-choice scenarios, and "score" for ordered rubrics 1.
Currently, the technology is integrated into several internal Google products to demonstrate its utility 1. For instance, Gemini Notebook utilizes the API for real-time dynamic model routing, which has reduced latency by 40%, while Search employs it as a safety layer for detecting MCP tools 1. Additionally, Stitch leverages the API for rapid design iterations 1. Beyond Google's ecosystem, the Gemini Decisions API has established integrations with major third-party platforms and frameworks, including OpenRouter, LangChain, Vercel, LiteLLM, and Pydantic AI 1.
评论
还没有评论,欢迎留下第一条。