Codex

• 1. overview (685 chars)

  OpenAI Codex is an independent programming agent tool first released by OpenAI in May 2025 and officially GA in October of the same year. After a large-scale upgrade in March 2026, it has evolved from a simple code generation tool to an AI agent system for the entire development process. It supports four usage forms: terminal CLI, desktop application, IDE extension and cloud, with parallel task processing, sandbox isolated execution and automatic test verification as its core competencies. The actual success rate of clearly defined tasks is about 74%. The monthly fee starts at US$20. The main competitors are Anthropic Claude Code and GitHub Copilot. The current biggest shortcomings are weak fuzzy instruction processing capabilities, limited access to private packages, and sandbox cold start delays.

• 2. Product background (1300 chars)

  Codex inherits its name from the code model API launched by OpenAI in 2021, but the two are completely different products. The earlier Codex API was retired in 2023 and the current version is a new autonomous programming agent based on the codex-1 model. On May 16, 2025, OpenAI CEO Sam Altman officially announced the launch of Codex, positioned as a "software engineering agent running in the cloud." In September of the same year, it was open sourced on GitHub in the form of CLI, built using Rust, and the protocol is Apache-2.0. It was officially released as GA in October and is fully open to ChatGPT Plus, Pro, Business and Enterprise users.   March 2026 is a critical node for Codex. OpenAI has released five major updates intensively this month: GPT-5.4 model integration, GPT-5.4 mini sub-agent model online, Codex App expanded to Windows platform, Plugin plug-in system and Automations automation functions, and Codex Security security review agent. These updates upgrade Codex from "helping you write code" to "helping you organize engineering work."   The developer's subscription system is embedded in the ChatGPT package and is not charged separately. Free users have very limited temporary access, the Plus package costs $20 per month, the Pro package costs $200 per month, and Business and Enterprise are customized plans. Codex adopts a points-based consumption model, and different models consume different amounts of points - GPT-5.4 consumes about 7 points per message (local tasks), and GPT-5.4 mini only consumes about 1 point.

• 3. Product functions and usage experience (2942 chars)

  The core workflow of Codex can be summarized into four steps: the user associates the GitHub repository and describes the task, Codex automatically clones the code base in an isolated environment and performs modifications, automatically runs the test suite for verification after completing the code changes, and finally presents a clean diff for the user to review. Users do not need to intervene in the entire process. Codex is like a "junior developer on the branch" - you assign work orders, it writes code, and you are responsible for code review.   Codex currently has established a "four-surface system", which is four main interfaces that work together. As a long-term core portal and agent "command center", Codex App is suitable for parallel task promotion and review control, and supports multi-agent collaboration visualization and Worktree isolated execution. CLI is suitable for senior developer workflows that are close to local repositories, scriptable, and rapid iteration. IDE extensions (supporting VS Code, Cursor, Windsurf) are suitable for the most contextual operations such as modification while writing, partial reconstruction, and single-module completion. Codex Cloud (web version) is suitable for clean task outsourcing and asynchronous execution, and finally recycles clean diff or PR.   Parallel processing is the core capability that differentiates Codex from most competing products. Users can initiate multiple independent tasks at the same time, and each task runs in an isolated sandbox environment without interfering with each other. A common usage scenario is to launch multiple low-priority maintenance tasks in the morning—document updates, style adjustments, simple bug fixes, etc.—and let Codex batch process them in the background while the developers themselves focus on higher-priority work.   AGENTS.md is a project-level configuration file introduced by Codex, which is equivalent to the project's "AI Guide". By placing AGENTS.md in the project root directory and each subdirectory, developers can tell Codex the coding specifications, architectural patterns, test requirements, and toolchain configuration of the project. According to OpenAI official tests, correctly configuring AGENTS.md can increase code generation accuracy from 40% to 75%, and code style consistency from 60% to 95%.The Automations feature added in March 2026 allows developers to turn frequently repetitive engineering chores into scheduled automated background tasks, such as CI failure summary, Release Brief generation, dependency upgrade scanning, and issue triage. Codex Security marked the product's transition from "generating code" to a high-trust process of "review, verification and patching". In the initial deployment, it scanned 1.2 million code submissions and identified 10,561 high-risk vulnerabilities.   Judging from the actual measurement data, Codex's performance on different types of tasks varies significantly. A reviewer tested 43 tasks within two weeks, and the results showed that the success rate of bug fixing was 75%, the test generation success rate was 80%, the reconstruction success rate was 75%, and the function implementation success rate was 63%, while the success rate of fuzzy open tasks was only 20%. Excluding ambiguous tasks, the overall success rate was 74%. Python and TypeScript are the top-performing languages ​​in Codex. Go has correct syntax but often misses idiomatic patterns.

• 4. Pricing and business model (1241 chars)

  Codex charges no separate subscription fee and is fully embedded within ChatGPT’s paid plans. Specifically, ChatGPT Plus has a monthly fee of US$20, can use Codex local tasks but has a limited message quota (about 33-168 messages within a 5-hour rolling window, affected by model selection), and does not support cloud tasks and code review. ChatGPT Pro costs $200 per month, has a higher message limit (approximately 223-1120 messages/5-hour window), and gets priority access. ChatGPT Team is $30 per user per month and has about 25 tasks per user per day. Business and Enterprise plans are priced customizable and support enterprise-level features such as cloud tasks and code review.   For heavy developers, the Plus package's limit of about 15 tasks per day may be exhausted before noon, and they need to evaluate whether to upgrade to Pro. API users pay by Token. It is estimated that the input is about 0.01-0.03 US dollars per 1K tokens, and the output is about 0.03-0.12 US dollars per 1K tokens. There is no usage limit.   In terms of cost performance, if you can save 5-10 hours of bug fixing and test writing time per month, the Plus package can pay for itself. The Pro package needs to replace about 1 day of development work per month to break even, and is only suitable for heavy users. The common suggestion from many reviewers is to "only subscribe with monthly payment, not annual payment" because AI tools iterate very quickly and maintaining flexibility is the first principle.

• 5. User reviews (1689 chars)

  Positive user reviews of Codex focus on several areas. The parallel task processing capability is generally considered to be the biggest highlight. Some developers commented that it realizes "multi-threaded workflow" and is suitable for people who manage dozens of warehouses. Bug fixing and test generation are recognized as the best scenarios. Some reviewers mentioned that Codex solved a FastAPI race condition problem that had been backlogged for weeks in 3 minutes, and discovered a date parsing bug hidden in the code when generating tests. In the 2026 Hengping Review, many developers believed that the GPT-5.4 model is more rigorous in change review than Claude Code's Opus 4.6 and will not make rash large-scale changes to the code. This is especially valuable in the reconstruction of large projects. The UI interface of Codex App is generally considered to be more user-friendly than Claude Code. The diff viewer has a GitHub PR-like experience and supports file-by-file approval or rejection.   Negative feedback focuses on a few pain points. Sandbox cold start delay is a frequent complaint - about 30 seconds for small warehouses and 60-90 seconds for large projects. The accumulated waiting time of frequent small tasks makes some users irritated. The processing ability of fuzzy instructions is weak. Open instructions such as "improve error handling" often lead to adding try-catch everywhere, with a success rate of only 20%. Iterating on the same branch for multiple rounds has a poor experience, and each iteration tends to create a new PR, which makes complex refactorings that require multiple steps very cumbersome. Private package access was a core impediment in earlier versions, and while the March 2026 update supports registry credential mounting, there are still limitations for enterprise environments where credentials cannot be configured. In April 2026, some domestic users reported that "large areas of accounts were banned". The specific reasons are still unclear.

• 6. Industry evaluation (1254 chars)

  Industry media and reviewers generally regard Codex as the most direct benchmark product of Claude Code. The positioning of the two is almost the same - an independent AI programming agent, supports Worktree, and can be deeply integrated into mainstream AI editors. The core difference lies in the underlying model: Codex uses OpenAI’s GPT series of models, and Claude Code uses Anthropic’s Claude model.   Multiple Hengping articles in 2026 show that Codex’s advantages lie in parallel processing, UI experience and model detail, while Claude Code is superior in depth of code understanding, context maintenance and execution speed. Many reviewers believe that the relationship between the two is not "choose one". In actual work, the more common combination is Copilot for real-time completion, Codex for batch autonomous tasks, and Claude Code for complex local reconstruction. Hengping directly recommends "Cursor + Codex" as one of the preferred combinations for professional developers.   The launch of Codex Security is seen as an important signal for OpenAI to enter the field of code security review. The data of 1.2 million code submission scans and the identification of 10,561 high-risk vulnerabilities, although still in the research preview stage, has demonstrated the feasibility of AI-driven security review in actual engineering. The addition of the Plugin plug-in system and Automations automation functions allows Codex to evolve from a single tool to a platform.

• 7. Controversies and Risks (1234 chars)

  Codex currently faces several major controversies and risks. In terms of security and privacy, although Codex adopts strict sandbox isolation and a two-stage trust boundary design (the Setup stage key is available and the main Agent stage key is automatically removed), many companies have concerns about handing over the code warehouse and API keys to third-party cloud services. The Codex CLI provides the disable_response_storage = true configuration option to disable OpenAI from storing conversation data, but in Cloud mode the data inevitably passes through the OpenAI server.   The "account ban" discussion that occurred in the domestic community in April 2026 deserves attention. A Zhihu article mentioned that the technical group was "whining", but the specific reason and scale of the ban have not yet been officially confirmed. This may involve issues with the usage specifications of the API transfer service, or it may be related to OpenAI’s cleanup actions against illegal accounts. For users who rely on Codex for daily development, this poses a service continuity risk.   The risk of dependency lock-in cannot be ignored either. Codex is deeply bound to OpenAI's models and ecosystem. Although the CLI version supports configuring third-party models (DeepSeek, Gemini, Ollama, etc.), the core experience and best performance still rely on the GPT series. If OpenAI adjusts its pricing or restriction policies, migration costs will not be cheap.

• 8. Applicable groups and usage suggestions (1170 chars)

  Codex is suitable for developers who can write clear and specific task descriptions. Explicit input like a Jira ticket ("Replace logger.warn() with logger.warning(), involving 23 files") is far better than vague instructions ("Improve error handling"). Python or TypeScript projects that require batch processing of test generation, bug fixes, and specific refactorings will see the best return on investment. Users who value the security isolation of their environment and are unwilling to run AI agents locally will also benefit from Codex's cloud sandbox architecture.   Scenarios that are less suitable for Codex include complex requirements that expect AI to understand architectural context and make design decisions, enterprise environments that rely heavily on private internal packages and cannot configure credentials, workflows that need to handle a large number of open-ended exploratory programming tasks, and real-time collaboration scenarios that cannot tolerate 30-90 second task startup delays.   For developers who are already using the OpenAI ecosystem (ChatGPT, GPT model API), Codex CLI has the lowest learning cost. For those using the Claude family of models and primarily doing deep local development, Claude Code may be the more natural choice. The best strategy for both is to try the monthly version and make a decision based on your actual workflow.

• 9. Summarize (829 chars)

  Since its release in May 2025, OpenAI Codex has developed from an experimental cloud programming agent to a complete AI programming system covering CLI, desktop applications, IDE and cloud. It performs well on bug fixing, test generation and well-defined refactoring tasks (about 74% success rate), and its parallel processing capabilities and safe sandbox isolation are its core competencies that differentiate it from competing products. However, weak fuzzy instruction processing, cold start delay and private package access restrictions are still pain points that need improvement. With the GPT-5.4 model upgrade in March 2026, the addition of the Plugin plug-in system and Automations automation functions, Codex is evolving from a tool that "helps you write code" to a platform that "helps you organize engineering work." For Python or TypeScript developers who can manage tasks with clear instructions, the $20/month Plus plan is a reasonably cost-effective option.

No in-depth report yet

• 

  SharonRussell_88

  —

  Codex 用了一周把原来三个下午才能搞完的第三方 API 对接全干完了，我就写了几句注释它就自动把序列化、重试、超时异常处理全封装好了，比自己手写还稳，连单元测试都顺手生成了，效率提升太明显了。

---

• 

  暖阳_25

  —

  跟 Copilot 最大的区别是它真能理解你下一步要做什么，不是补全代码而是预判意图，用久了会上瘾。

---

• 

  Harold.BaileyK

  —

  云端多任务并行是真的爽，一次性丢三个 bug 给它修，它自己开三个沙箱跑，我这边该干嘛干嘛，效率直接拉满。

---

• 

  若梦790

  —

  最大惊喜是它帮我从打字员变回了思考者。以前 70% 的时间写业务逻辑处理脏活累活，现在反过来，大部分时间在思考架构和边界条件，能力重心从怎么写彻底迁移到了怎么想，这个转变真的价值连城。

---

• 

  JulieBqiley

  —

  后端工程化能力比 Claude Code 强不少，干活细致，但跑得确实慢，适合睡前丢任务第二天验收。

---

• 

  LAcas

  —

  太强了！

---

• 

  Jason.ScottZ

  —

  我们团队已经把它纳入日常工具链了，QA 和开发都在用。给清晰的需求加测试覆盖率，Codex 产出的东西基本可以直接合，省去了大量来回 review 的时间，确实解放了生产力，开发幸福感提升了不止一个档次。

---

• 

  Grace.Young_20248

  —

  省钱了省钱了！走 API 调用一次才几美分，比 Cursor 的月订阅划算多了，适合个人开发者。

---

• 

  WDiaz_2021

  —

  刚上手的时候觉得这玩意儿也就那样，用了一周就真香了。修 bug 和写测试特别靠谱，基本不用怎么改就能直接用。

---

• 

  JerryBrown_2024

  —

  重构老项目的测试架构，Codex 帮了大忙。它把遗留的旧测试全部识别出来然后批量重写重构成新框架，又自动化又规范，虽然还得人工 review 一遍但省了 80% 的工作量，团队整体效率提升非常明显，是今年引入最有价值的工具。

---

• 

  Christopher.Robinson520

  —

  Agent 风格的工作流确实好用，一个指令下去它自己规划步骤、识别文件、跨文件改代码、跑测试验证，最后出 PR 带 citation，透明度和可控性都很好，比我想象中成熟多了，团队 review 起来也更轻松了。

---

• 

  骑士602

  —

  刚开始用还挺满意的，但新版的 AGENTS.md 配置真的把可定制性拉满了，把团队编码规范、架构风格丢进去它就能按统一的风格来写代码，这一点 Copilot 现在还完全做不到，对维护大项目的团队来说太重要了。

---

• 

  NodeRunner736

  —

  话说免费的额度够用吗？我团队四个人在试用，感觉企业版费用和实际产出比还算划算，但个人开发者用的话确实要考虑成本问题，有没有用过的老哥说说经验？

---

• 

  SatoshiSnake153

  —

  最让我惊艳的是它居然会自己判断场景决定用流式输出还是同步模式，我以为这种底层细节得我写 Prompt 来回引导才能实现，结果 Codex 自动选了最合适的方案，真的聪明，感觉是有经验的搭档而不是一个工具。

---

• 

  ChristinaHicks_7

  —

  用了四个月了，整体体验比预期好，不是那种花里胡哨的新奇玩具，是真能解决实际问题的生产力工具。建议搭配强测试覆盖率使用效果最佳，给的任务越清晰它输出的质量越高，这个投入产出比相当可观，值得推荐给每个开发团队。

---

• 

  silverbear317

  —

  yyds。

---

• 

  BrittanyKing_X35

  —

  Codex 这波更新太猛了，背景计算机控制直接接管 macOS 应用，Claude 封号限流之后果断转过来了。

---

• 

  狗狗_12

  —

  用了快一年了，从最初的 40-60% 成功率到现在 85-90%，稳定性提升太明显。

---

• 

  清风_1

  —

  建议加上手机遥控功能，这样更方便。

---

• 

  tiflyrn

  —

  用了两天，感觉还行吧，查错确实牛，速度能接受，不过偶尔会陷入自循环，希望官方能优化一下这个问题。

---

• 

  Mark.Jenkins_Max

  —

  天下苦 A 社久矣，Codex 这波接住了。

---

• 

  LUm_or

  —

  说实话，有点失望，本来以为能完全替代人工编程，结果遇到复杂逻辑还是要自己来，期望过高了。

---

• 

  Aaron2422024

  —

  说实话，Codex 的 Tab 补全和 Claude Code 比还是差点意思，但胜在和 ChatGPT 生态打通。

---

• 

  天涯_14

  —

  Codex的PR审查功能太强了，团队之前漏掉的bug都能发现，这钱花得值了。

---

• 

  jwle8scq

  —

  多智能体并行真的香，同时跑三个任务一点都不卡，效率直接翻倍。

---

• 

  trueLavanyaChiplunkar_dev

  —

  沙箱模式太赞了，再也不用担心AI乱改本地文件，安全感拉满。

---

• 

  smy2tt09g

  —

  macOS 用户狂喜，背景计算机控制终于来了！

---

• 

  DWhite_77552

  —

  作为在 WorkOS Applied AI 团队工作的开发者，用了 Codex 快一年了。最明显的感受是稳定性提升巨大，从最初的 40-60% 成功率到现在 85-90%，而且任务失败时提供的错误信息清晰多了，不像以前那样让人摸不着头脑。多轮对话也更可靠了，可以就实现细节来回讨论，不用每次都启动新任务。

---

• 

  DRoss520

  —

  用了两天，内存泄露 bug 还挺烦人的，而且会话只能存档不能删除，有点难受。

---

• 

  k2s5_l8

  —

  刚更新的桌面应用控制功能很实用，能直接操作Mac应用了，自动化工作流终于不是梦。

No reviews yet