Skip to content

Google Launches Gemini Decisions API for Scoring Up to 1,000 Questions

Baoyu · XRelayed· @dotey·· 1 day agoSame news

2.6K views · 9 likes · 2 reposts (at collection)

Original
Who
Google
What
introduced Gemini Decision 1, a frontier decision model in the Gemini API
Quote
Google 今天在 Gemini API 里上线了 Decisions API,配套新模型 Gemini Decision 1。

What happened

Google launched the Gemini Decisions API with its Gemini Decision 1 model, which returns answers and confidence scores instead of generated text. A request can include up to 1,000 questions about text, images, audio, video, or PDFs; input costs $0.10 per million tokens, with no charge for output tokens. The post compares its limits, features, and Google-reported benchmark results with OpenAI's Decisions API and TypeSafe's Jev 1.14.

The same story is shown in the picks as 《Google Introduces Gemini Decision 1 for Scoring 1,000 Questions in One Pass》

Read next

See the full storyGoogle Launches Gemini Decisions API for Batch Question Scoring

Reported by 6 sources

    Body · English edition pending (original content)

    Google 推出 Gemini Decisions API:不写回答只做判断,一次最多问 1000 个问题,输入每百万 Token 0.1 美元、输出不收钱

    Google 今天在 Gemini API 里上线了 Decisions API,配套新模型 Gemini Decision 1。它不生成文字,只答题:给它一份输入和一组问题,它直接返回每道题的答案,外加一个“有多大把握”的置信度。

    【它解决什么问题】

    做 AI 应用时,很多调用其实只要一个判断:这条客服消息转给哪个组?这个问题交给便宜模型还是推理模型?这张图违不违规?以前只能让大模型写一段话,再从里面解析答案,慢,容易解析失败,还要为输出 Token 付钱。

    Decisions API 只有三种题型:是非题,返回“是”的概率;单选题,2 到 255 个选项,返回选中项和每个选项的概率;打分题,按你写的档位打分。每个答案带 0 到 1 的置信度,代码可以设门槛,比如高于 0.7 自动执行,低于就转人工。官方说这些概率是校准过的,意思是模型说 90% 的时候,大约十次里真对九次,所以置信度能直接拿来当门槛。

    Google 博客的例子:上传一张茶水间台面的照片,同时问三个问题。“有没有洒了的液体”答 0.95;“怎么处理”选“用抹布擦”,置信度 0.98;“脏乱程度按 0 到 2 打分”给 1.19,置信度 0.77。

    【多快、多便宜】

    一次请求的输入最多 256K Token,文字、JSON、图片、音频、视频、PDF 可以混着放,比如一段客服通话录音加账户信息。同一份输入上最多挂 1000 个问题,彼此互不影响。输入只计一次费,每多一个问题只多算这道题本身的字数。

    Google 称输入在 1.6 万 Token 以内时,端到端延迟 150 到 250 毫秒,问 1 个和问 100 个问题耗时差不多;同一批评测下,比用 Gemini 3.5 Flash-Lite 生成答案便宜近 5 倍。Gemini Notebook 已经用它给每个提问挑模型,日常问题延迟最多降了 40%。

    【谁能用】

    在 Google AI Studio 拿个 API Key 就能调,模型名 gemini-decision-1,走 v1beta 接口。有免费层,每分钟 200 次、每天 1 万次,免费层的数据会被 Google 用来改进产品,付费层不会。OpenRouter、LangChain、Vercel 等正在接入,Google Cloud 上的 Gemini 要稍后才支持。

    不支持上下文缓存和 Batch API,同一份大文档反复问,每次都按全价算输入;不能直接传 YouTube 链接;问题本身只能写文字。

    【跟 OpenAI 的同名产品比】

    OpenAI 早一个月做了同样的东西,名字也叫 Decisions API:10 月 6 日出测试版,10 月 22 日正式开放,用 GPT-6 Luna,同样每百万输入 Token 0.1 美元、输出不收费。区别是 OpenAI 一次最多 400 个问题,只收文字、图片和 JSON,不收音视频和 PDF;但它支持缓存,重复输入能打折,还能直接塞进对话历史和工具调用记录。

    创业公司 TypeSafe 的 Jev 1.14 也专做这类判断,在 OpenRouter 上每百万输入 Token 0.042 美元,只收文字和图片。

    Google 博客附的对比表里,Gemini Decision 1 在 9 项测试中 7 项第一(其中音频、视频两项只有它和 Flash-Lite 参测),校准误差 0.025,GPT-6 Luna 是 0.072;网页操作和“该不该调用工具”两项输给了 Jev 1.14。7 项共同评测的总花费,Jev 最低,0.68 美元,Gemini Decision 1 是 0.88 美元,GPT-6 Luna 0.93 美元。以上跑分都是 Google 自己测的。

    Google 博客:blog.google/innovation-and-ai…
    官方文档:aistudio.google.com/docs/deci…
    OpenAI Decisions API 文档:developers.openai.com/api/doc…

    Quoted postSundar Pichai@sundarpichai

    Today we’re launching the Gemini Decisions API, powered by our new Gemini Decision 1 model in @GoogleAIStudio.

    Instead of generating text token by token, it’s built for closed-set choices scoring up to 1,000 questions across text, images, audio, video + PDFs in a single pass. A few details:

    • 3 decision types: yes/no, multiple choice (up to 255 options) + rubric scoring
    • $0.10 / 1M input tokens with $0 billed for output tokens (~5x cheaper than Flash-Lite)
    • leads on several benchmark evals with lowest calibration error

    Google teams are already building with it, below is a look at Stitch Rapid designing a full product flow in seconds (1x speed)

    View the quoted post on X

    The story so far

    Google launched the Gemini Decisions API, powered by Gemini Decision 1, for scoring up to 1,000 questions in one request. It returns answers with confidence scores for yes-or-no, multiple-choice questions with up to 255 options, and rubric scoring across text, images, audio, video, and PDFs. The API costs $0.10 per million input tokens with no output-token charge; Google says it also has a free tier with limits of 200 requests per minute and 10,000 per day. Google says free-tier data may be used to improve its products, while paid-tier data will not. The API is available through Google AI Studio. Google Cloud support is planned for later. It does not support context caching or the Batch API, and users cannot submit YouTube links directly.

    How the story developed · 2

    View full event
    1. 11/12 07:09 (UTC+8)

      Google Launches Gemini Decisions API for Scoring Up to 1,000 Questions This report

    2. 11/12 04:22 (UTC+8)
      Google Introduces Gemini Decision 1 for Scoring 1,000 Questions in One PassThe Decoder