跳到内容
现在

科技 · Gemini Decisions API

Google 推出可在不撰写答案的情况下作出决策的新 Gemini 模型

Gemini Decision 1 返回概率而非散文式文本，每次请求最多可就共享材料提出 1,000 个问题。它是否有用，取决于这些快速判断是否足以支持实际行动。

文字 阅读需 7 分钟 · 20 个来源

HeyDex

原语言视频

译自英文原文。

当 AI 模型不再撰写答案，而是开始提供让软件采取行动的判断时，会发生什么？Google 的 Gemini Decisions API 将这个问题落到了实际层面：它承诺能对文档、图像、录音及其他材料快速作出有界判断，且不生成任何文本 token。其优势在于减少生成散文式文本的时间，而这些文本还需要由另一个程序来解读。风险则是把包装整齐的概率误当成可靠的裁决。⁠2⁠3

Google DeepMind 于 2026 年 11 月 11 日推出该 API，公告由产品经理 Min Choi 和首席科学家 Felix Yu 发布。Google 表示，该 API 由 Gemini Decision 1 驱动，可在一次请求中，就共享材料评估最多 1,000 个彼此独立的判断题、选择题和评分标准题。这是供其他软件调用的接口，而不是另一种聊天机器人界面。⁠2

这一区别很重要。传统生成式模型会逐个 token 构建答案；需要作出决策的应用程序随后可能还得从答案中提取所需内容。Gemini Decision 1 则在一次处理过程中评估有界选项，并返回结构化判断和概率。Google 的文档称，软件可以直接根据响应进行分支处理、排序和路由，输出 token 用量始终为零。它不会提供开放式解释。⁠1⁠3

0生成的文本 token 3
最多 1,000宣传的每次请求可针对共享输入提出的问题数 2
$0.10据报每百万输入 token 的价格 1

因此，只要所需答案是给周边应用程序的一条指令，该端点就可能派上用场：选择模型、检查政策、排列相关材料，或决定请求应转到何处。其工程优势显而易见。如果软件需要的是有界判断，让模型先写一段话再处理，就是可以避免的绕路。⁠1⁠3

这些判断实际包含什么

Hybrid.co.id 的发布报道介绍了开发者可在一次调用中混合使用的三种问题类型。判断题用于估计某个条件是否成立，并可选填何种情况算作真或假。选择题接受 2 到 255 个候选项，并返回在这些选项之间归一化的概率。评分标准题使用从零开始排列的等级，返回连续加权分数，以及各个等级对应的概率。⁠1

这些定义将相当大的责任交给开发者。仍然需要有人选择备选项、定义评分等级，并决定接下来采取什么行动。格式正确的响应无法弥补问题设定不当。选择概率最高的选项，也不能证明应用程序一开始就提供了正确的选项。

可以看发布报道中提到的保险示例：理赔申请 PDF 可与损坏照片一同提交。应用程序可以就这些材料提出有界问题，并利用概率决定下一步。这是合理的受理机制，但并不能证明模型能够安全地批准或拒绝理赔。仅仅因为两者都可以表示为分类任务，就模糊了文件路由与决定某人是否享有权益之间的区别。⁠1

宣传的输入类型包括文本、图像、音频、视频和 PDF，可在同一次请求中组合使用。Hybrid 报告称，上下文窗口最多为 256,000 个 token（即系统处理的单位），并且可调整图像或视频分辨率。降低分辨率可以减少 token 消耗，但会损失视觉细节。例如，客户通话录音可以与账户信息一同提交，而不必作为完全独立的输入处理。⁠1

所谓可处理 1,000 个问题，也需要准确理解其含义：这是针对共享材料提出许多问题，并非以零额外成本处理 1,000 份互不相关、完整规模的输入材料。每个问题都带有自己的指令和标准。⁠1

快速且便宜，但有边界

Hybrid 报告称，对于最多 16,000 个 token 的上下文，模型计算时间中位数为 50–100 毫秒，端到端延迟为 150–250 毫秒。报告还称，当问题数量从 1 个增至 100 个时，响应时长几乎保持稳定。这些说法的范围比宣传的最高容量更窄：它们并未证明在 1,000 个问题或最大宣传上下文长度下也能达到同样速度。⁠1

目前有关此次发布的报道中，速度数据尚未得到独立复现。也没有可复现的说明交代硬件、服务区域、并发量、预热条件或响应时间的完整分布。应用开发者应把这些数据视为开展测试的理由，而非服务级别保证。⁠20

据报道，价格为每百万输入 token $0.10，不收取输出 token 费用。据称，在一次请求中，共享材料只计费一次，无论有多少问题使用这些材料；但问题的指令和标准仍会增加输入 token。Hybrid 还表示，不收取可变的思考 token 费用，也不另收音频转录和光学字符识别费用。因此，该服务并非免费处理材料。零输出费用这一点比折扣更根本：没有生成的答案文本可供计费。⁠1⁠3

一项发布对比称，在等效基准测试中，通过 Interactions API 调用该端点的价格“便宜近五倍”，相比对象是 Gemini 3.5 Flash-Lite。若没有可复现的方法，这就不是普遍适用的节省幅度。实际成本取决于提交的材料、问题的数量和长度，以及它所取代的工作负载。⁠1

置信度仍需安全网

Google 将返回的概率描述为经过校准：大致而言，这意味着预测的可能性应当与实际正确的频率相符。Hybrid 表示，系统会提供 0 到 1 范围内的置信度信息，开发者可据此设置阈值，以启动备用流程或人工审核。高数值仍然只是预测，并非保证。⁠1⁠3

发布材料宣传了该模型在路由、推理、排序和多模态理解方面的表现，并声称校准误差最低。但材料并未提供经过独立验证的校准测量结果、总体错误率或可复现的准确率对比。这些缺失在错误分类会触发难以撤销的行动时尤其重要。⁠1⁠20

Google 所描述的内部用途展现了预期角色。名为 Gemini Notebook 的工作流会选择模型层级和思考级别，对较复杂的查询才采用更深入的推理；据称，该工作流处理日常查询的延迟最多可降低 40 percent。这项说法仅适用于该工作流，并非适用于所有使用此端点的应用。在 Google Labs 的 CC 应用中，任务更简单：在不到一秒的时间内，从 200 多个选项中挑选一个表情符号。⁠1

其他示例也需要区分决策模型与其周边产品。据称，Stitch 在涉及交互式线框图以及语音或触控修改的快速画布实验中使用该模型。Google Search 的一个示例则检查模型上下文协议工具——连接 AI 系统与工具的接口——是否存在对抗性行为和有害内容。这个安全示例没有给出检测准确率或误报率的数值。它展示的是预期用途，而非已证实的攻击免疫能力。⁠1

据称，用户可通过 Google AI Studio 中的 Gemini API 使用该服务，并需提供 API 密钥；平台还提供 Python、JavaScript 和 REST 指南，以及免费的原型开发层级。模型标识符为 gemini-decision-1。公布的生态支持包括 google-genai 软件开发套件，以及与 OpenRouter、LangChain、Vercel 和其他框架的集成。这些是发布时的支持声明，并非经核实的客户生产环境结果。据称，Google Cloud 支持即将推出，但尚无确切日期。⁠1⁠2

发布报道尚未明确哪些国家符合使用资格，也未说明该服务处于预览版还是正式可用阶段。更重要的考验仍在后头：在开发者自己的工作负载上，速度、成本和置信度质量能否保持。永远不写答案的模型可能很有价值，但软件仍需要充分理由，才能根据模型隐含的答案采取行动。

来源19

在此处引用的来源中，Google最先发布，比本报道播出早42小时。 随后还有13家媒体发布。
  1. Google
  2. AlphaSignal
  3. AI Unfiltered with Thorsten Meyer
  4. Adiust
  5. Aiholics
  6. Writeble
  7. Walter Pinem
  8. LLMs Blog
  9. 另有+6家媒体
  10. HeyDay News

5家媒体未注明日期：Google DocsAppFor.aidatacamp.comPapers with Codeventurebeat.com

按媒体查看所有来源

来自 19 家媒体的 20 个来源· 引用最多的优先

播出原貌
  1. Google's new decision model generates zero text tokens.
  2. Google advertises up to 1,000 questions about shared material in one request. The results can steer software, check policies or grade documents, while processing the input still costs money.
  3. The practical question is how reliably those fast judgments work when software acts on them.
  4. An insurance workflow could combine a claim PDF with photographs of damage. Software could then ask bounded questions about that material and use the returned probabilities to decide what happens next.
  5. Google DeepMind introduced the Gemini Decisions API on November 11, 2026. Gemini Decision 1 powers it. An API is a connection that lets other software use the model directly.
  6. Google says one request can evaluate up to 1,000 independent questions about the supplied material. Software can use the results directly to branch, sort or route work, without first extracting a decision from a written answer.
  7. Gemini Decision 1 evaluates candidate answers in a single processing pass. A chatbot normally builds its answer piece by piece. This endpoint returns structured judgments and probabilities, so a program doesn't need a sentence to parse.
  8. Developers can mix three question types in one call. Yes-or-no questions estimate whether a condition holds. Multiple-choice questions accept two to 255 candidates. Rubric questions assess material against ordered scoring levels.
  9. Developers can describe what counts as true or false. Choice questions return probabilities normalized across the alternatives. Rubrics begin at zero and return a continuous weighted score, alongside probabilities for individual levels. These outputs provide a bounded judgment rather than an open-ended explanation.
  10. A customer-call recording can travel with account information in the same input. The advertised formats also include text, images, video and PDFs. Developers can combine media instead of submitting every item separately.
  11. The advertised context window holds up to 256,000 tokens, the units the system processes. Developers can reduce image or video resolution to use fewer tokens, trading away visual detail. Each added question also brings its own instructions and criteria.
  12. The advertised timing figures cover contexts up to 16,000 tokens. Median model computation is 50 to 100 milliseconds. The reported end-to-end range is 150 to 250 milliseconds, including more than the model's computation alone.
  13. Response duration is described as staying nearly steady between one and 100 questions. Those figures don't promise the same speed at 1,000 questions or the largest context. An application's workload and serving conditions still shape the wait.
  14. Subscribe to this channel to follow whether these speed and confidence claims hold up in real applications.
  15. The launch price is reportedly ten cents per million input tokens, with zero dollars charged for output tokens. The endpoint generates no text tokens. Shared input is billed once within a request, even when several questions use it.
  16. Every question's instructions and criteria still add input tokens. The pricing description excludes variable thinking-token charges and separate fees for audio transcription or reading text from images. A no-cost prototyping tier is also described.
  17. A launch comparison describes the endpoint as almost five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API. That claim comes from equivalent benchmark testing; the saving for a particular application depends on its actual workload.
  18. Google also promotes results in routing, reasoning, ranking and multimodal understanding, including a lowest-calibration-error claim. Calibration describes how closely predicted probabilities match actual outcomes. The speed, cost and calibration claims haven't been independently reproduced, so developers still need to test their own traffic.
  19. A workflow called Gemini Notebook uses the model to select model tiers and thinking levels. It carries a claimed latency reduction of up to 40 percent for everyday queries, reserving deeper reasoning for more complex work.
  20. That improvement belongs to the named workflow. Stitch is described as using the endpoint in rapid canvas experiments with interactive product-flow wireframes and voice- or touch-driven revisions. The surrounding application produces the experience; the endpoint supplies decisions.
  21. CC, a Google Labs application, uses it to choose an emoji from more than 200 options in a fraction of a second. The examples show the intended role as a quick control layer without training a separate classifier.
  22. Another described use checks Model Context Protocol tools in Google Search for adversarial behavior and harmful content. That protocol connects AI systems with tools. The safety example provides no numerical detection accuracy or false-positive rate.
  23. Confidence information ranges from zero to one, allowing developers to set thresholds for fallback procedures or human review. Google calls its probabilities calibrated. A high score still doesn't guarantee that a classification is right.
  24. Developers choose which alternatives exist and what each rubric means. They also decide when an uncertain or consequential result goes to a person. In a claims workflow, a fast classification alone can't prove the model understood the underlying documents correctly.
  25. Access is described through the Gemini API in Google AI Studio using an API key. Python, JavaScript and REST guides accompany demonstrations named Doorstep Parcels, Speed Sketch and Find the Zoo Keeper.
  26. The model identifier is gemini-decision-1. Developers still need clear access terms for their location and workload. Country eligibility, quotas and whether the offering is in preview or general availability haven't been settled in the launch details.
  27. Announced support includes the google-genai software kit and integrations involving OpenRouter, LangChain and Vercel. The list also includes LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. These are launch integration claims, rather than demonstrated customer production results.
  28. Availability through Gemini on Google Cloud is described as forthcoming, without a firm date.

译文更新于

近期播出