OpenAI 使决定 API 与 GPT-6 Luna 通用

AI 模型和平台

OpenAI 使决定 API 与 GPT-6 Luna 通用

在 Google 上的首选来源中添加 Unite.AI

OpenAI将其决定API在2026年10月22日普遍公布，宣布公司将GPT-6 Luna 的动力端点从公共β中取出，价格为每百万个输入令牌0. API 返回应用程序用来对内容，路由请求进行分类和排列工作优先次序的有约束，打字的答案(概率，选择，或标点分数).

OpenAI 称 API将Luna的智力集中在有限定答案的问题上，使得实时的决定能够从文本，图像，以及结构化的输入中实现。 公司称端点通过反应API提供这些答案的速度可达10x.

2026年9月22日，OpenAI 推出月球与GPT-6 Sol并列，月球定价为每百万输入令牌0.1美元，每百万输出令牌0.5美元。 称公司将两个型号的API价格与他们的GPT-5.6促销定价相比削减了50%，并称对称推进了成本效益前沿；它继续将GPT-6 Astra描述为其全局最佳模式。 在API中，两种型号为gpt-6-sol和gpt-6-luna.

API的决定如何运作

根据OpenAI的开发者文档，请求转到POST/v1/Decisions,gpt-6-luna是目前唯一支持的模型。 每个请求对一组共享输入(文本字符串或命令的一系列消息，工具呼叫，工具输出)，在1到400个问题之间，答复包含一个响应每个问题开发者指定名称的答案阵列。

API支持三个问题类型。 上游问题返回一个概率从0到1，一个条件是真实的，比如产品照片中的可见损害。 一个选择问题选择开发者预先定义的选项之一，并返回每个选项的概率和一个置信值。 一个分数问题将输入值对准2到10个按序的数值级，并返回该等级指数的概率加权平均值，因此一个分数可以在各等级之间下降。

开发者可以通过用户信息中的输入 json内容部分直接传递结构化数据，图像可以作为内置基底64数据URL或作为可公开获取的HTTP或HTTPS URL提供。 由于对话历史可以作为输入提供，API可以结合函数和自定义工具调用及其输出来评价一个请求。 文档称文件，音频和项目引用不支持，端点不接受文件 id输入。

对于依赖于更早答案的决定，文档指示开发者分别发送请求。 它还将构建器指向在响应API中调用的结构化输出或函数，当它们需要自定义的JSON对象或模型请求的工具调用而不是三个输入的答案中的一个。

自公开竞选以来的增补

OpenAI 称在公众beta后增加了原生的GPT-Live融合，快速缓冲再利用上下文，以及在美国和欧盟的推论居住能力，再加上亚洲几个地区的附加推论能力。 开发者现在可以通过本地的JSON输入，将图像包括在问题指令中，并且每个请求最多可询问400个问题，是之前限制的两倍。

文件的新清单还包括托管的HTTP和HTTPS图像输入、有工具呼叫和产出的对话-历史评价、组织使用情况报告和私人链接支助。 使用端点 GET /v1/organization/usage/ decisions，需要一个带有 api.usage 的管理员 API 密钥。 读取权限并返回令牌使用量和请求数，它们可以被工程、API密钥或模型过滤或分组。

在快速缓存下，重复请求可以重复使用一个匹配的共享输入前缀至少长1,024个令牌，并获得折扣的缓存读取定价；质疑内容，包括特定问题图像，坐落在可重复使用的前缀之外。 API默认在共享输入的结尾放置一个隐性缓存断点，开发者可以使用调用器提供的即时 cache key来分割再利用，或者切换到明确标记的断点。

通过GPT-Live集成，一个语音应用程序可以在输入或输出笔录到达时运行决定，Live发布一个决定事件供应用程序采取行动。 兹Q 1Z 实例配置，添加到会话中。 触发，从笔录更新中选择avatar的面部表情，公司描述一个客户支持agent使用相同的模式来决定是问一个澄清问题，路由一个请求，还是将一个问题升级。

报告的使用情况和早期评价

OpenAI 称自己的团队已经在产品和业务工作流程中应用API：点球团队用它来缩短响应时间和改善互动，个人财务团队用它来对交易描述进行分类，通信团队用它来按照主题对覆盖进行分类，并得分覆盖是否强化或挑战关键叙述，帮助确定哪些内容值得更仔细的审查。

公告中引述的客户在自己的评价中报告了速度和成本增益。 Box负责平台和集成产品管理的副总裁Biswaroop Palit、称和Box Extract 联合决定API，以与Resolution API相同的准确性，在30个保险证书案件中提供了大约5x更快的路线选择，向审查者发送了信心较低的案例。 Datadog 员工工程师 AJ Stuyvenberg 称 API 在执行前对 agent 生成的rshell 命令的风险进行评分，比现有的边缘模型快40x p95 耐用，精确度相当。 Ramp研究工程师René Sultan 称 API将人工智能成本归因于基础工程，比使用GPT-6 Luna更快，更便宜30%，不牺牲准确性。

Sierra工程师Ajeet Grewal 称，API在Sierra的现场客户对话中，在不到280毫秒的p95中选择下一项动作，将GPT-Live中的决定API集成时间缩短至下一个动作

定价和数据控制

决定API每百万个输入令牌花费0.10美元，每百万个缓存输入令牌花费0.1美元，OpenAI收费仅用于输入令牌，没有输出卡通收费。 区域加工溢价和长文本输入定价乘数适用，缓存阅读率和缓存写率符合标准gpt-6-luna快速缓存率。

美国和欧洲支持区域处理(欧洲经济区加瑞士)，所有受支持的API区域都有数据居住权，符合资格的客户可以使用零数据保留和HIPAA. API可以通过专用的决定端点和OpenAI平台上的决定游戏场访问，开发者可以在写代码之前根据自己的文本或图像测试问题。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

AI Models & Platforms

OpenAI Makes Decisions API Generally Available with GPT-6 Luna

Add Unite.AI to your preferred sources on Google

OpenAI made its Decisions API generally available on October 22, 2026, the company announced, taking the GPT-6 Luna–powered endpoint out of public beta at a price of $0.10 per million input tokens. The API returns bounded, typed answers (a probability, a selection, or a rubric score) that applications use to classify content, route requests, and prioritize work.

OpenAI said the API focuses Luna’s intelligence on questions with bounded answers, enabling near real-time decisions from text, images, and structured inputs. The company said the endpoint delivers those answers up to 10x faster than asking GPT-6 Luna through the Responses API.

OpenAI introduced GPT-6 Luna alongside GPT-6 Sol on September 22, 2026, pricing Luna at $0.10 per million input tokens and $0.50 per million output tokens. The company said it cut API prices for both models by 50% compared with their GPT-5.6 promotional pricing and described the pair as advancing the frontier on cost efficiency; it continues to describe GPT-6 Astra as its best model across the board. In the API, the two models are available as gpt-6-sol and gpt-6-luna.

How the Decisions API Works

Requests go to POST /v1/decisions, and gpt-6-luna is currently the only supported model, according to OpenAI’s developer documentation. Each request pairs shared input (a text string or an ordered array of messages, tool calls, and tool outputs) with between 1 and 400 questions, and the response contains an answers array that echoes each question’s developer-assigned name.

The API supports three question types. A predicate question returns a probability from 0 to 1 that a condition is true, such as visible damage in a product photo. A choice question selects one of the developer’s predefined options and also returns probabilities for each option and a confidence value. A score question rates an input against two to ten ordered rubric levels and returns the probability-weighted average of the level indices, so a score can fall between levels.

Developers can pass structured data directly through an input_json content part in a user message, and images can be supplied as inline base64 data URLs or as publicly accessible HTTP or HTTPS URLs. Because conversation history can be supplied as input, the API can evaluate a request in context alongside function and custom tool calls and their outputs. The documentation states that files, audio, and item references are not supported, and the endpoint does not accept file_id inputs.

For decisions that depend on an earlier answer, the documentation directs developers to send separate requests. It also points builders to Structured Outputs or function calling in the Responses API when they need a custom JSON object or a model-requested tool call rather than one of the three typed answers.

Additions Since the Public Beta

OpenAI said that after the public beta it added a native GPT-Live integration, prompt caching for reused context, and inference residency in the United States and EU, plus additional inference capacity in several regions in Asia. Developers can now pass native JSON inputs, include images in question instructions, and ask up to 400 questions per request, twice the previous limit.

The documentation’s what’s-new list further includes hosted HTTP and HTTPS image inputs, conversation-history evaluation with tool calls and outputs, organization usage reporting, and Private Link support. The usage endpoint, GET /v1/organization/usage/decisions, requires an admin API key with the api.usage.read permission and returns token usage and request counts that can be filtered or grouped by project, API key, or model.

Under prompt caching, repeated requests can reuse a matching prefix of shared input at least 1,024 tokens long and receive discounted cache-read pricing; question content, including question-specific images, sits outside the reusable prefix. The API places an implicit cache breakpoint at the end of shared input by default, and developers can partition reuse with a caller-provided prompt_cache_key or switch to explicitly marked breakpoints.

With the GPT-Live integration, a voice application can run decisions as input or output transcripts arrive, and Live emits a decision event for the application to act on. OpenAI’s example configuration, added to session.triggers, selects an avatar’s facial expression from transcript updates, and the company describes a customer support agent using the same pattern to decide whether to ask a clarifying question, route a request, or escalate an issue.

Reported Use and Early Evaluations

OpenAI said its own teams already apply the API across product and business workflows: the dots team uses it to reduce response times and improve interactivity, the Personal Finance team uses it to categorize transaction descriptions, and the Communications team uses it to classify coverage by theme and score whether the coverage reinforces or challenges key narratives, helping identify what warrants closer review.

Customers quoted in the announcement reported speed and cost gains in their own evaluations. Box’s vice president of product management for platform and integrations, Biswaroop Palit, said the Decisions API combined with Box Extract delivered roughly 5x faster route selection across 30 certificate-of-insurance cases at the same accuracy as the Responses API, sending lower-confidence cases to reviewers. Datadog staff engineer AJ Stuyvenberg said the API scored the risk of agent-generated rshell commands before execution, delivering 40x faster p95 latency than available frontier models with comparable accuracy. Ramp research engineer René Sultan said the API made attributing AI costs to the underlying work 4x faster and 30% cheaper than using GPT-6 Luna, without sacrificing accuracy.

Sierra engineer Ajeet Grewal said the API chooses the next action during parts of Sierra’s orchestration in live customer conversations at under 280 milliseconds p95, about 3x faster and 90% lower cost per call than the company’s previous approach, with no drop in customer outcomes. Speak CTO Andrew Hsu said the Decisions API integration in GPT-Live reduced time to next action by more than 90% in the company’s testing, from 1.4 seconds to 0.13 seconds.

Pricing and Data Controls

The Decisions API costs $0.10 per million input tokens and $0.01 per million cached input tokens, and OpenAI charges only for input tokens, with no output-token charges. Regional processing premiums and long-context input pricing multipliers apply, and cache-read and cache-write rates match the standard gpt-6-luna prompt caching rates.

Regional processing is supported in the United States and Europe (the EEA plus Switzerland), data residency is available in all supported API regions, and Zero Data Retention and HIPAA use are available to eligible customers. The API is accessible through the dedicated Decisions endpoint and a Decisions Playground on OpenAI’s platform, where developers can test questions against their own text or images before writing code.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.