OpenAI

Vendor profile — OpenAI. Reviewed 2026-10-18 from vendor documentation only; no hands-on testing.

厂商档案 — OpenAI。于 2026-10-18 依据厂商文档复核；没有实际操作测试。

Profile概况

Organization机构
OpenAI
Category类别
Frontier lab that sells models through an API and agent products through ChatGPT and Codex (interpretation: it is a model and agent runtime vendor, not an MLOps platform vendor).前沿模型实验室：通过 API 销售模型，通过 ChatGPT 与 Codex 提供智能体产品（解读：它是模型与智能体运行时的供应商，不是 MLOps 平台供应商）。
Primary buyer主要买家
Two audiences appear in the docs: developers and engineering teams (API platform), and workspace owners and admins (ChatGPT Enterprise and Edu, Compliance Platform) [1][6]. This is an observation from who the documentation addresses, not a sales statement.文档面向两类读者：开发者与工程团队（API 平台），以及工作区所有者与管理员（ChatGPT Enterprise 与 Edu、合规平台）[1][6]。这是对文档读者的观察，不是销售口径。
Relevant products相关产品
Responses API [13]; Agents SDK and Traces dashboard [14][4]; Datasets, Evals, graders and trace grading (being retired) [2][8][3]; Agent Builder and ChatKit [5][3]; Batch API [12]; fine-tuning (being wound down) [10][3]; Codex [7]; ChatGPT Compliance Platform [1]; dots [15].Responses API [13]；Agents SDK 与 Traces 仪表板 [14][4]；Datasets、Evals、graders 与 trace grading（正在退役）[2][8][3]；Agent Builder 与 ChatKit [5][3]；Batch API [12]；微调（正在收缩）[10][3]；Codex [7]；ChatGPT 合规平台 [1]；dots [15]。
Last reviewed最后复核
2026-10-18

MDLC coverageMDLC 阶段覆盖

Coverage by stage按阶段的覆盖情况 — ● full完整 · ◐ partial部分 · ○ none无 · ? not yet researched尚未研究
Stage阶段Coverage覆盖Product / feature产品 / 功能Source来源
Require需求?No first-party requirements or intake feature found in the documents read. Only a survey of documentation titles; absence is not proven, so the stage stays “?”.所读文档中没有发现厂商自有的需求/立项功能。只是浏览了文档标题，不能证明不存在，因此保持“?”。—
Plan规划?No first-party planning feature (model choice, tool set, autonomy level) found beyond model documentation. Not researched in depth.除模型文档外，没有发现厂商自有的规划功能（模型选择、工具集、自主程度）。未深入研究。—
Build构建◐Responses API with built-in web search, file search, computer use, code interpreter and remote MCP [13]. Agents SDK: agents, handoffs, guardrails, sessions, human-in-the-loop [14]. Codex on CLI, IDE extension, cloud and ChatGPT apps [7]. Retreating: self-serve fine-tuning (SFT, DPO, RFT) is closed to organisations that never ran it since 2026-05-07, and Agent Builder shuts down 2026-11-30 [3][10][5]. Interpretation: the build tooling that remains is code-first.Responses API 内置网页搜索、文件搜索、计算机使用、代码解释器与远程 MCP [13]。Agents SDK：agents、handoffs、guardrails、sessions、人在回路 [14]。Codex 覆盖 CLI、IDE 扩展、云端与 ChatGPT 应用 [7]。正在收缩：自 2026-05-07 起，从未微调过的组织不能再创建自助微调任务；Agent Builder 将于 2026-11-30 关闭 [3][10][5]。解读：保留下来的构建工具以代码为主。[13][14][7][3][10][5]
Validate验证◐Datasets, Evals, graders (scores 0 to 1) [2], trace grading [8] and a dataset-backed prompt optimizer [11]. OpenAI states it is deprecating the Evals platform: read-only 2026-10-31, shut down 2026-11-30, with no OpenAI replacement listed; the deprecations page points to a Promptfoo migration path [2][3] (summariser output, lower confidence). The Moderation API is free [9]. Interpretation: first-party evaluation is retiring within weeks of this review.Datasets、Evals、graders（0 到 1 分）[2]，trace grading [8]，以及依赖数据集的提示词优化器 [11]。OpenAI 声明正在弃用 Evals 平台：2026-10-31 变为只读，2026-11-30 关闭，未列出 OpenAI 自己的替代品；弃用页指向迁移到 Promptfoo [2][3]（摘要器输出，可信度较低）。Moderation API 免费 [9]。解读：本次复核后数周内，厂商自有的评测将退役。[2][8][11][3][9]
Approve审批◐Human approval of agent actions, not of a release: for dots, users set per-action handling (act without asking, act when told, ask first, hand off); changing a password or moving money needs the human to take over, and some actions need approval each time [16]; Enterprise dots access is gated by admin switches [15]. Codex documents agent approvals, permission profiles and Auto-review [7] (titles only). No release sign-off or launch-review workflow for a customer’s model or app was found.针对智能体动作的人工批准，而非发布批准：对 dots，用户可按动作类型设置处理方式（直接做、被告知再做、先询问、交还给人）；改密码或转账必须由人接手，有些动作每次都需批准 [16]；Enterprise 的 dots 访问由管理员开关控制 [15]。Codex 文档列有智能体审批、权限配置与 Auto-review [7]（只读了标题）。没有发现针对客户模型或应用的发布签批/上线评审流程。[16][15][7]
Deploy部署◐Hosted API for OpenAI models (Responses API; Assistants API shut down 2026-08-26) [13][3]. Batch API at a 50% discount with a 24-hour window [12]. Per-project data residency in ten regions, in-region processing only for US, EU and UAE [6]. Agent Builder publishes a versioned workflow that can be embedded with ChatKit or exported as Agents SDK code; Agent Builder is retiring, ChatKit stays [5][3]. Interpretation: serving covers OpenAI’s own models, not a customer’s own model artefacts.托管 OpenAI 模型的 API（Responses API；Assistants API 已于 2026-08-26 关闭）[13][3]。Batch API 五折，24 小时窗口 [12]。按项目配置数据驻留，十个区域，仅美国、欧盟、阿联酋支持区域内处理 [6]。Agent Builder 发布带版本的工作流，可用 ChatKit 嵌入或导出为 Agents SDK 代码；Agent Builder 即将退役，ChatKit 保留 [5][3]。解读：服务范围是 OpenAI 自己的模型，不是客户自有模型制品。[13][3][12][6][5]
Monitor监控◐Traces dashboard for Agents SDK runs: generations, tool calls, handoffs, guardrails. Tracing is on by default, unavailable under Zero Data Retention, and the docs list about 35 community and vendor processors (for example Datadog, Langfuse, MLflow) [4]. Documentation navigation lists workspace analytics, Usage Insights and an Analytics API for Codex [7] (titles only; content not read). No production quality or drift monitoring product found.Agents SDK 运行的 Traces 仪表板：生成、工具调用、handoff、guardrail。追踪默认开启，零数据保留（ZDR）下不可用；文档列出约 35 个社区/厂商处理器（如 Datadog、Langfuse、MLflow）[4]。文档导航列出工作区分析、Usage Insights 与 Codex 的 Analytics API [7]（只读了标题，未读内容）。没有发现生产环境质量或漂移监控产品。[4][7]
Governance治理◐ChatGPT Enterprise and Edu Compliance Platform: immutable, append-only compliance logs kept 30 days, plus a stateful API, for eDiscovery, DLP and SIEM; listed partners include Microsoft Purview, CrowdStrike, Palo Alto Networks and Zscaler [1]. API data controls: Zero Data Retention or Modified Abuse Monitoring per organisation or project, data residency, Enterprise Key Management [6]. Enterprise dots access, local computer access and custom rules are off by default [15]. Codex docs list Governance, Agent security and a Compliance API page [7] (titles only).ChatGPT Enterprise 与 Edu 合规平台：不可变、仅追加的合规日志（保留 30 天）及状态型 API，对接电子取证、DLP 与 SIEM；列出的合作方包括 Microsoft Purview、CrowdStrike、Palo Alto Networks、Zscaler [1]。API 数据控制：按组织或项目配置零数据保留或修改版滥用监控、数据驻留、企业密钥管理 [6]。Enterprise 的 dots 访问、本地电脑访问与自定义规则默认关闭 [15]。Codex 文档列有治理、智能体安全与合规 API 页面 [7]（只读了标题）。[1][6][15][7]

Observations观察

OpenAI’s own documents say it is retiring lifecycle tools: the Evals platform (read-only 2026-10-31, shut down 2026-11-30) and Agent Builder (2026-11-30) within about six weeks of this review, and self-serve fine-tuning in stages through 2027-01-06; the Assistants API was shut down on 2026-08-26. [3][2][5][10]

OpenAI 自己的文档说，正在退役生命周期工具：Evals 平台（2026-10-31 只读，2026-11-30 关闭）与 Agent Builder（2026-11-30）在本次复核后约六周内结束，自助微调分阶段收缩至 2027-01-06；Assistants API 已于 2026-08-26 关闭。[3][2][5][10]

The Compliance Platform page lists Microsoft Purview among the partners that ingest updated conversation logs, and Microsoft’s Purview page lists ChatGPT Enterprise as a supported enterprise AI app. [1] See the Microsoft profile.

合规平台页面把 Microsoft Purview 列为接收新版对话日志的合作方；Microsoft 的 Purview 页面也把 ChatGPT Enterprise 列为受支持的企业 AI 应用。[1] 见 Microsoft 页面。

Agents SDK tracing is on by default and records inputs and outputs unless trace_include_sensitive_data is turned off (default true); under Zero Data Retention tracing is unavailable. [4]

Agents SDK 追踪默认开启，除非关闭 trace_include_sensitive_data（默认 true），否则会记录输入输出；在零数据保留下追踪不可用。[4]

Method and limits: this review read vendor documentation only, no hands-on testing. Most OpenAI pages were read through a page summariser, not verbatim (marked in References); only the Compliance Platform help page was read in full. Navigation-only evidence for Codex administration is flagged “titles only”.

方法与局限：本次只读厂商文档，没有实际操作。多数 OpenAI 页面是通过页面摘要器读取的，不是逐字阅读（见参考来源标注）；只有合规平台帮助页是完整阅读。Codex 管理侧仅有导航标题的证据，已标注“只读了标题”。

Interpretation解读

OpenAI supplies the model and the agent runtime, and is thinning out its own lifecycle tooling (evaluation, visual agent builder, fine-tuning) while keeping the code-first path (Responses API, Agents SDK, Codex) and the enterprise audit export. That fits the idea that build and evaluation tooling stays specialised while control-plane functions live elsewhere (H4 on the thesis page), but the evidence is one vendor’s deprecation notices, so it is weak support.

OpenAI 提供模型和智能体运行时，同时在削减自己的生命周期工具（评测、可视化智能体构建、微调），保留代码优先路径（Responses API、Agents SDK、Codex）和企业审计导出。这与“构建与评测工具保持专门化，控制平面功能落在别处”的想法（论点页的 H4）一致，但证据只是单一厂商的弃用通知，支持力度弱。

The Compliance Platform is an export interface for other tools (SIEM, DLP, eDiscovery), not a governance system of record. For an MDLC buyer, the register of models, approvals and evidence would have to live outside OpenAI.

合规平台是面向其他工具（SIEM、DLP、电子取证）的导出接口，而不是治理的记录系统。对 MDLC 的买家来说，模型登记、审批和证据需要放在 OpenAI 之外。

Strategic position战略位置

FACT. Deprecation dates and replacements are as stated on the vendor’s deprecations page [3].

FACT。弃用日期和替代方案以厂商弃用页为准 [3]。

INTERPRETATION. Strengths: frontier models, a widely used code-first agent stack, and an audit export that security vendors already ingest. Gaps: no requirements, planning or release-approval tooling found; first-party evaluation is leaving; monitoring is tracing only.

INTERPRETATION。优势：前沿模型、被广泛使用的代码优先智能体栈、安全厂商已经接入的审计导出。缺口：没有发现需求、规划或发布审批工具；厂商自有评测正在退出；监控只有追踪。

HYPOTHESIS. OpenAI will lean on partners (evaluation, observability, governance vendors) for lifecycle functions it retires, rather than rebuilding them. Evidence that would test this: what replaces Evals after 2026-11-30, and whether partner lists grow.

HYPOTHESIS。OpenAI 退役的生命周期功能会交给合作伙伴（评测、可观测、治理厂商），而不是自己重建。检验证据：2026-11-30 之后 Evals 由什么替代，合作方名单是否增加。

Open questions待解问题

What replaces OpenAI Evals after 2026-11-30, and is Promptfoo an official path or only a pointer? (Check the deprecations page again after the date.)

2026-11-30 之后 OpenAI Evals 由什么替代？Promptfoo 是官方路径还是仅仅一个指引？（日期过后重查弃用页。）

What does the Compliance Platform actually export for Codex and dots actions (event schemas)? The Admin API reference was not read.

合规平台对 Codex 与 dots 的动作究竟导出什么（事件模式）？Admin API 参考文档尚未阅读。

Do Codex permission profiles, Auto-review and agent approvals (titles only so far) amount to an Approve stage for agent autonomy? Read those pages in full.

Codex 的权限配置、Auto-review 与智能体审批（目前只读了标题）是否构成针对智能体自主程度的“审批”阶段？需完整阅读这些页面。

Require and Plan are “?”: is anything there beyond model documentation?

Require 与 Plan 仍为“?”：除模型文档外是否还有别的？

References参考来源

  1. OpenAI Help Center, “OpenAI Compliance Platform for Enterprise and Edu customers,” read in full 2026-10-18. https://help.openai.com/en/articles/9261474-openai-compliance-platform-for-enterprise-and-edu-customers vendor documentation · S-0061
  2. OpenAI API docs, “Getting started with datasets” (Datasets and Evals; Evals deprecation notice). https://developers.openai.com/api/docs/guides/evaluation-getting-started vendor documentation (via summariser) · S-0062
  3. OpenAI API docs, “Deprecations.” https://developers.openai.com/api/docs/deprecations vendor documentation (via summariser) · S-0063
  4. OpenAI Agents SDK (Python), “Tracing.” https://openai.github.io/openai-agents-python/tracing/ vendor documentation (via summariser) · S-0064
  5. OpenAI API docs, “Agent Builder.” https://developers.openai.com/api/docs/guides/agent-builder vendor documentation (via summariser) · S-0065
  6. OpenAI API docs, “Your data” (retention, ZDR, MAM, data residency, EKM). https://developers.openai.com/api/docs/guides/your-data vendor documentation (via summariser) · S-0066
  7. OpenAI, ChatGPT and Codex documentation hub (navigation and landing page only). https://learn.chatgpt.com/docs vendor documentation (via summariser) · S-0067
  8. OpenAI API docs, “Trace grading.” https://developers.openai.com/api/docs/guides/trace-grading vendor documentation (via summariser) · S-0068
  9. OpenAI API docs, “Safety best practices.” https://developers.openai.com/api/docs/guides/safety-best-practices vendor documentation (via summariser) · S-0069
  10. OpenAI API docs, “Model optimization” (fine-tuning methods and wind-down notice). https://developers.openai.com/api/docs/guides/model-optimization vendor documentation (via summariser) · S-0070
  11. OpenAI API docs, “Prompt optimizer.” https://developers.openai.com/api/docs/guides/prompt-optimizer vendor documentation (via summariser) · S-0071
  12. OpenAI API docs, “Batch API.” https://developers.openai.com/api/docs/guides/batch vendor documentation (via summariser) · S-0072
  13. OpenAI API docs, “Migrate to the Responses API.” https://developers.openai.com/api/docs/guides/migrate-to-responses vendor documentation (via summariser) · S-0073
  14. OpenAI Agents SDK (Python), overview. https://openai.github.io/openai-agents-python/ vendor documentation (via summariser) · S-0074
  15. OpenAI Help Center, “Manage dots in ChatGPT workspaces” (already registered; see the dots note). https://help.openai.com/en/articles/20001554-manage-dots-in-chatgpt-workspaces vendor documentation · S-0019
  16. OpenAI Help Center, “Dots privacy, security, and safety FAQs” (already registered; see the dots note). https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs vendor documentation · S-0020