Google发布Gemini Decision 1决策API
Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens
Agent开发者的新利器，专为多路分支决策设计的专用API，低延迟且零输出计费，值得在复杂路由链路中压测对比。
Google has launched the Gemini Decisions API, powered by a new model, Gemini Decision 1 (gemini-decision-1). The Gemini Decisions API does not generate text. It scores up to 1,000 typed questions against one shared input and returns calibrated probabilities. For developers, this targets the small branching calls inside agents: routing, policy checks, tool selection and grading. Google prices it at $0.10 per 1M input tokens, with zero output tokens billed.
Google 推出了由新模型 Gemini Decision 1（gemini-decision-1）驱动的 Gemini Decisions API。Gemini Decisions API 不生成文本。它针对共享输入对多达 1,000 个键入问题进行评分，并返回校准后的概率。对于开发者而言，这旨在处理智能体内部的少量分支调用：路由、策略检查、工具选择和评分。Google 的定价为每 1M 输入令牌 0.10 美元，输出令牌不计费。
TL;DR
- Size: Context is 262,144 tokens (256K) across shared input and all questions.
- Runs on: Google-hosted API only (Gemini Developer API, v1beta). No open weights or self-hosting.
- Performance: Leads 7 of 9 accuracy benchmarks in Google’s table, with the lowest calibration error.
- Best: 92.2% on JevBench (public split).
- Worst: 58.7% Macro-F1 on Amazon ESCI, its lowest score. It also trails Jev 1.14 on When2Call (79.9% vs 82.0%).
- Bottom line: Best: multimodal, calibrated decisions in 150-250 ms. Worst: not the cheapest; Jev 1.14 costs $0.68 vs $0.88 on the shared suite.
- 容量：共享输入和所有问题的上下文总计为 262,144 个令牌（256K）。
- 运行平台：仅限 Google 托管 API（Gemini Developer API，v1beta）。无开源权重或自托管选项。
- 性能：在 Google 表格中的 9 项准确率基准测试中领先 7 项，且校准误差最低。
- 最佳表现：JevBench（公开分割集）上达到 92.2%。
- 最差表现：Amazon ESCI 上的 Macro-F1 为 58.7%，这是其最低分。它在 When2Call 上也落后于 Jev 1.14（79.9% vs 82.0%）。
- 总结：最佳表现：150-250 毫秒内提供多模态校准决策。最差表现：并非最便宜；Jev 1.14 在共享套件中的成本为 0.68 美元，而 Gemini Decision 1 为 0.88 美元。
What is the Gemini Decisions API?
什么是 Gemini Decisions API？
Gemini Decision 1 is a specialized multimodal model from Google built for closed-set decisions. It reads text, images, video, audio and PDFs in one prefill pass. Output token usage is always 0. Your code branches directly on the returned numbers.
Gemini Decision 1 是 Google 专为封闭集决策构建的专业化多模态模型。它通过一次预填充阶段读取文本、图像、视频、音频和 PDF。输出令牌使用量始终为 0。您的代码直接根据返回的数字进行分支判断。
Each request has 3 parts: a model ID, a shared input, and a map of questions. The docs define 3 question types:
每个请求包含 3 个部分：一个模型 ID、一个共享输入以及一个问题映射表。文档定义了 3 种问题类型：
- noul (yes/no): returns the probability of yes, bounded between 0.01 and 0.99.
- choice: picks 1 of 2 to 255 options and returns normalized probabilities for every option.
- score: rates input on 2 to 255 ordered rubric levels and returns a probability-weighted score.
- noul（是/否）：返回“是”的概率，范围限制在 0.01 到 0.99 之间。
- choice（选择）：从 2 到 255 个选项中选择一个，并为每个选项返回归一化概率。
- score（评分）：在 2 到 255 个有序的量表等级上对输入进行评级，并返回加权概率分数。
Every answer carries a confidence value from 0 to 1. Google suggests routing low-confidence cases to a fallback or a human reviewer. Questions are scored in parallel and cannot see each other.
每个答案都带有 0 到 1 之间的置信度值。Google 建议将低置信度的案例路由到备用方案或人工审核员。问题是并行评分的，且彼此不可见。
How does Gemini Decision 1 work?
Gemini Decision 1 是如何工作的？
Standard models decode an answer token by token, then your code parses it. Gemini Decision 1 instead scores your candidate options in one forward pass. Google reports median model compute of 50-100 ms and 150-250 ms end to end. That holds for inputs up to 16K tokens and stays nearly flat from 1 to 100 questions.
标准模型逐令牌解码答案，然后由您的代码进行解析。而 Gemini Decision 1 则在一次前向传递中对候选选项进行评分。Google 报告的中位模型计算时间为 50-100 毫秒，端到端时间为 150-250 毫秒。这对于高达 16K 令牌的输入均适用，并且从 1 个问题到 100 个问题几乎保持不变。
Billing follows the same design. The shared input is billed once per request, however many questions you ask. Each question adds only its own instruction and criteria tokens. There are no thinking-token, transcription or OCR charges.
计费遵循相同的设计。无论您提出多少问题，共享输入在每个请求中仅计费一次。每个问题仅增加其自身的指令和标准令牌。没有思考令牌、转录或 OCR 费用。
Gemini Decision 1 benchmarks
Gemini Decision 1 基准测试
Google tested Gemini Decision 1 against 3 rivals. They are TypeSafe Jev 1.14, GPT-6 Luna on OpenAI Decisions and Gemini 3.5 Flash-Lite. All figures below are self-reported by Google. Refusals and safety blocks count as incorrect.
Google 对 Gemini Decision 1 与 3 个竞争对手进行了测试。它们分别是 TypeSafe Jev 1.14、OpenAI Decisions 上的 GPT-6 Luna 以及 Gemini 3.5 Flash-Lite。以下所有数据均由 Google 自行报告。拒绝回答和安全拦截均计为错误。
| Benchmark | Gemini Decision 1 | TypeSafe Jev 1.14 | GPT-6 Luna (OpenAI Decisions) | Gemini 3.5 Flash-Lite |
|---|---|---|---|---|
| JevBench (public split) | 92.2% | 90.0% | 90.9% | 86.6% |
| Amazon ESCI (Macro-F1) | 58.7% | 56.7% | 54.7% | 47.2% |
| New Yorker Caption (text-only) | 79.7% | 72.0% | 68.4% | 73.7% |
| Mind2Web Multimodal | 86.2% | 86.8% | 81.8% | 78.0% |
| PubMedQA | 79.8% | 78.2% | 74.4% | 77.6% |
| Global-MMLU-Lite (non-English) | 86.0% | 85.0% | 79.6% | 78.8% |
| When2Call | 79.9% | 82.0% | 70.4% | 60.9% |
| MMAU (audio) | 71.8% | Not reported | Not reported | 67.7% |
| Video-MME (no subtitles) | 69.1% | Not reported | Not reported | 63.3% |
| Top-1 ECE (lower is better) | 0.0250 | 0.0494 | 0.0721 | 0.2104 |
| Total price, 7 shared evals | $0.88 | $0.68 | $0.93 | $4.14 |
| 基准测试 | Gemini Decision 1 | TypeSafe Jev 1.14 | GPT-6 Luna (OpenAI Decisions) | Gemini 3.5 Flash-Lite |
|---|---|---|---|---|
| JevBench（公开分割集） | 92.2% | 90.0% | 90.9% | 86.6% |
| Amazon ESCI（宏平均 F1 分数） | 58.7% | 56.7% | 54.7% | 47.2% |
| New Yorker Caption（纯文本） | 79.7% | 72.0% | 68.4% | 73.7% |
| Mind2Web Multimodal | 86.2% | 86.8% | 81.8% | 78.0% |
| PubMedQA | 79.8% | 78.2% | 74.4% | 77.6% |
| Global-MMLU-Lite（非英语） | 86.0% | 85.0% | 79.6% | 78.8% |
| When2Call | 79.9% | 82.0% | 70.4% | 60.9% |
| MMAU（音频） | 71.8% | 未报告 | 未报告 | 67.7% |
| Video-MME（无字幕） | 69.1% | 未报告 | 未报告 | 63.3% |
| Top-1 ECE（越低越好） | 0.0250 | 0.0494 | 0.0721 | 0.2104 |
| 总价格，7 项共享评估 | $0.88 | $0.68 | $0.93 | $4.14 |
Source: Google blog benchmark table.
来源：Google 博客基准测试表。
The calibration result is the clearest gap. Gemini Decision 1’s error of 0.0250 is about half of Jev 1.14’s 0.0494. That matters when you set confidence thresholds for automation. On cost, Google says it is nearly 5x cheaper than Flash-Lite across the shared suite ($0.88 vs $4.14).
校准结果是差距最明显的指标。Gemini Decision 1 的错误率为 0.0250，约为 Jev 1.14 的 0.0494 的一半。这在设置自动化置信度阈值时至关重要。在成本方面，Google 表示其在该共享套件中的费用比 Flash-Lite 便宜近 5 倍（$0.88 对比 $4.14）。
Gemini Decisions API vs OpenAI Decisions vs TypeSafe Jev
Gemini Decisions API 与 OpenAI Decisions 与 TypeSafe Jev
| Feature | Gemini Decision 1 | GPT-6 Luna, OpenAI Decisions | TypeSafe Jev 1.14 |
|---|---|---|---|
| Parameters | Not disclosed | Not disclosed | Not disclosed |
| Context | 262,144 tokens | 1,050,000 (model page) | 64K per request, 262K with opt-in header |
| Questions per request | 1 to 1,000 | 1 to 400 | No fixed cap, bounded by context |
| Input modalities | Text, image, audio, video, PDF | Text, image, JSON | Text (docs) |
| Input price per 1M tokens | $0.10 | $0.10 uncached | $0.042 |
| Output price | $0 | $0 | $0 |
| Prompt caching | Not supported | Supported, charged at Luna rates | Not disclosed |
| Latency claim | 150-250 ms end to end | About 10x faster than Responses API | Not disclosed |
| 功能 | Gemini Decision 1 | GPT-6 Luna, OpenAI Decisions | TypeSafe Jev 1.14 |
|---|---|---|---|
| 参数量 | 未披露 | 未披露 | 未披露 |
| 上下文窗口 | 262,144 tokens | 1,050,000（模型页面） | 每次请求 64K，通过 opt-in 标头可扩展至 262K |
| 每次请求的问题数 | 1 到 1,000 | 1 到 400 | 无固定上限，受上下文限制 |
| 输入模态 | 文本、图像、音频、视频、PDF | 文本、图像、JSON | 文本（文档） |
| 每百万 tokens 输入价格 | $0.10 | $0.10（未缓存） | $0.042 |
| 输出价格 | $0 | $0 | $0 |
| 提示词缓存 | 不支持 | 支持，按 Luna 费率收费 | 未披露 |
| 延迟声明 | 端到端 150-250 毫秒 | 比 Responses API 快约 10 倍 | 未披露 |
How to deploy the Gemini Decisions API
如何部署 Gemini Decisions API
The endpoint is live on the Gemini Developer API (v1beta) and needs google-genai SDK 3.2 or later. A no-cost tier is available for prototyping in Google AI Studio. Python, JavaScript and REST guides are in the documentation.
该端点已在 Gemini Developer API (v1beta) 上线，需要 google-genai SDK 3.2 或更高版本。Google AI Studio 提供用于原型设计的免费层级。Python、JavaScript 和 REST 指南可在文档中找到。
Ecosystem support is rolling out across OpenRouter, LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Support in Gemini on Google Cloud is listed as coming soon.
生态系统支持正在 OpenRouter、LangChain、Vercel AI SDK、LiteLLM、Pydantic AI、Pipecat、LiveKit 和 OpenCode 上逐步推出。Google Cloud 上的 Gemini 支持被列为即将推出。
Google already uses it internally. Gemini Notebook routes queries to model tiers with it, serving everyday queries with up to 40% lower latency. Google Labs’ CC picks emoji reactions from more than 200 options. Stitch’s rapid canvas uses it to build playable wireframes in seconds.
Google 已在内部使用它。Gemini Notebook 借助它将查询路由至模型层级，以高达 40% 的更低延迟处理日常查询。Google Labs 的 CC 从超过 200 个选项中选取表情反应。Stitch 的快速画布利用它在数秒内构建可交互的线框图。
Key Takeaways
关键要点
- Gemini Decision 1 scores up to 1,000 questions per call with zero output tokens.
- It reads text, images, audio, video and PDFs natively in 1 pass.
- Calibration error of 0.0250 is half that of TypeSafe Jev 1.14.
- At $0.10/M input tokens, it still costs more than Jev’s $0.042/M.
- No caching or Batch API yet, so repeated context is billed in full.
- Gemini Decision 1 每次调用可评分多达 1,000 个问题，且输出 token 为零。
- 它原生通过一次传递读取文本、图像、音频、视频和 PDF 文件。
- 校准误差为 0.0250，仅为 TypeSafe Jev 1.14 的一半。
- 按每百万输入 token 0.10 美元计算，其成本仍高于 Jev 的每百万 0.042 美元。
- 目前尚无缓存或 Batch API，因此重复上下文将全额计费。
Check out the technical details, docs and model page. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
查看技术细节、文档和模型页面。此外，欢迎在 Twitter 上关注我们，别忘了加入我们拥有 15 万+成员的 ML SubReddit 并订阅我们的新闻通讯。等等！你在 Telegram 上吗？现在你也可以加入我们的 Telegram 群组。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力