# Codex explained as a working agent, with every account path routed through a storefront > A Chinese tutorial repository that defines Codex as a programming agent, compares the GPT-5.6 model tiers, and prices Fast mode, then spends its longest section on subscription shops and third-party relays. The explanation is the useful half. The acquisition tables are the half to read carefully. **xianyu110/gpt-codex** — 这是一个关于 OpenAI CodeX 的完整教程网站，专为国内开发者打造。 - Repository: https://github.com/xianyu110/gpt-codex - Website: https://codex.maynorai.top/ - Stars: 1,221 · Forks: 91 - Language: HTML - License: not declared - Published: 2026-09-15 · Updated: 2026-09-15 · Language: en - Canonical page: https://hysenlabs.com/projects/xianyu110-gpt-codex ## Codex is defined as an agent over a project directory, not a chat window The tutorial opens by fixing the category. Codex is presented as OpenAI's answer to Claude Code, and the category it claims is a programming agent rather than a chat product. The difference it draws is not about intelligence but about behaviour: it behaves like a workflow tool, it keeps pushing a task forward, and its purpose is to move something along instead of answering one prompt. A three row table does the comparison work. Vendor, OpenAI against Anthropic. Underlying model, GPT-5.6 Sol, Terra and Luna against Claude Opus 4.6. Positioning, a programming agent trending toward a general agent against a programming agent. Then the argument for why that category matters. Several decades of digitised work settled into code, software, interfaces, automation, system configuration and data pipelines, so an assistant that handles those artefacts can do far more than write code. The stated conclusion is that coding agents are turning into a general purpose interface to work, and that their endpoint is approaching a general agent. ## The model name moved once, and the citation still points at the old version The tutorial spends a full warning on an identification it considers obsolete: Codex is not simply GPT-5.3 codex. The earlier generation is characterised as a programming specialist, a pure coding worker. Its replacement, announced on 2026-04-23, is quoted from OpenAI as the newest flagship model and the strongest agentic coding model available. The summary offered is a shift in role, from worker to engineering lead: able to plan, execute and check its own work, and better suited to long context and multi stage tasks. A four row table sets the old belief against the new one across model identity, style, suitable work, and what the change means for someone who does not write code. The reference list is where the care slips. Three official OpenAI references are offered, but two of them are named with no address at all, and the one that is linked ends in the slug introducing-gpt-5-5. That is not the version the section is about, so the single working citation on the page points at a different model release than the one it is used to support. ## Fast mode buys about 1.5 times the speed for about 2.5 times the token cost Two numbers carry the practical weight of the model update, and one of them is a cost. Fast mode is described as running roughly 1.5 times faster, at roughly 2.5 times the token cost, and is recommended for tasks where waiting is the problem. That is the whole trade stated in one line, and it is the only place in the tutorial where the new model is given a price rather than a description. The other number is the context window, given as 400K tokens on the Codex side, where the listed account tiers extend beyond the consumer plans. The plan table itself is deliberately hedged. Plus, Pro, Business and Enterprise can use the GPT-5.6 models in both ChatGPT and Codex. Edu and Go appear on the Codex availability list. The free tier is not listed, and the page's instruction for that case, and for every case, is to check the model list inside your own account. It also states that availability changes with region, plan and rollout pace, and that only what was published on the release date is written down. ## Four backup rows in the relay table, three destinations The third party relay section is presented as a table, and the table has a defect worth noticing before you rely on it. The rows are a GPT service homepage, a Gemini service homepage, a domestic Codex login page, four rows labelled as backup addresses, a one month bundle combining Codex and image generation, a backup shop link, and a row for configuring a relay API through a separate tool. The four backup rows point at three places. The first is a login path on one domain, and the second, third and fourth all name the same bare domain with no path at all. The shop link then appears twice: once as the primary purchase row in this table, and again as the backup row in the tool table further down. So the fallback ladder is thinner than it looks. If the main entry point is unstable, three of the four rungs are the same address, and the label that claims four alternatives describes two. For a reader who is already on an unstable network, that is a specific gap rather than a general one. ## Every account path ends at a shop, and the page says so in one line Three acquisition routes are documented, and only one of them is an OpenAI purchase. The first is an official subscription, reached through a 20 dollar top-up link with a second shop offered as backup, and the route is marked as requiring the reader to supply their own network workaround. The second is the relay scheme, and this is the part to read carefully: the page states in a single note that the bundle and the relay configuration are third party relays and not an official OpenAI subscription, while offering that they need no network workaround and suit people who want it working quickly. The third runs through a configuration tool described further down. What the note does not cover is what the relay means in practice. Every prompt, file and credential you use travels through an endpoint that the page does not name, and there is no operator, retention policy or expiry date attached to any of those entries. The fastest documented route to Codex in this repository is also the one that hands a third party your account and your code. ## The relay configuration claims a daily quota and runs through someone else's tool One row of the relay table describes configuring a relay API alongside a separate configuration switcher, and claims a daily allowance of 150 dollars. That figure belongs to the seller's offer, not to OpenAI, and nothing on the page explains how the allowance is calculated or what happens when it runs out. The tool is the third acquisition route, given as three steps: download and install the switcher from its own GitHub releases page, configure the relay API against it, then read a full tutorial article hosted on the author's news site. The tool is not written by this repository's author; it lives in a separate project, and this tutorial links only its releases page and leaves the install to you. The stated audience for this route is people who would rather follow a finished tutorial than debug an integration from zero, and people who care more about it working quickly than about the original path. The cost is that a configuration switcher plus a relay endpoint means your accounts are configured in software that neither project here describes. ## The install instructions that exist are for Mac, and there are two channels The client is treated as a desktop application, and the download path is short. OpenAI's Codex page at chatgpt.com/codex carries a download application button. On Mac, the page points at the App Store, where searching for Codex and installing from the store is described as the simplest route, and it also gives the official disk image at persistent.oaistatic.com/codex-app-prod/Codex.dmg. After that you install and sign in. So one platform gets two channels with different consequences. The store route hands updates to the store and needs no manual file handling. The disk image is a direct download from OpenAI's asset host, which means you are the one keeping the file current. The page treats the desktop app as an ordinary end user install rather than something to compile, and the repository itself publishes no GitHub releases, so there is nothing here to install from. The repository is the tutorial; the client comes from OpenAI. ## A static site whose printed update date trails its last push by three months The repository is a website, and its own layout explains how it is maintained. At the root sit the site entry page, a long Markdown tutorial file, a directory of documents, an English tree, a light variant, recipes, reference material, a favicon, a CNAME file that binds the site to codex.maynorai.top, a site configuration file, a URL mapping file, and a small Python script that uploads the screenshots. The primary language is HTML and the default branch is master. No GitHub releases have been published, so the repository has no versioned distribution of its own. The page prints an update date of 2026-04-27 and an author name, while the last push to master landed on 2026-08-20, so the printed date understates how recently anything here changed. A navigation block lists eight sibling tutorial repositories, and one of them is an external launcher for the Codex app that injects extra behaviour through the DevTools protocol while leaving the installed files untouched. ## Conclusion Read this repository for its explanation of what an agent style coding tool changes, and treat its acquisition section as a commercial listing rather than setup guidance. Three things to check before you act on any of it: read your own account's model list instead of the plan table, confirm the model name on OpenAI's own pages rather than through the link the tutorial provides, and remember that the relay route sends your prompts, your code and your credentials through an endpoint that nobody in this repository names. If you only need the concepts, the model comparison table and the Fast mode cost note are the two parts that age gracefully. ## FAQ ### What is Codex used for? The tutorial positions it as a programming agent that works around a project directory, files and a task goal rather than answering prompts in a chat window. The work it lists is code generation, engineering changes, moving a project forward, running commands, and breaking a technical task into steps, and it claims the category is trending toward a general agent. ### Is ChatGPT Codex free? The plan table lists Plus, Pro, Business and Enterprise as able to use the GPT-5.6 Sol, Terra and Luna models in both ChatGPT and Codex, adds Edu and Go on the Codex side, and does not list the free tier. Its instruction is to check the model list inside your own account, since availability changes with region and plan. No OpenAI price is stated on the page; the 20 dollar figure belongs to a third party top-up link. ### How do I install gpt codex? OpenAI's Codex page at chatgpt.com/codex carries a download application button. On Mac the tutorial points at the App Store listing for Codex as the simplest route, and also gives the official disk image at persistent.oaistatic.com/codex-app-prod/Codex.dmg, after which you install and sign in. ### How do I install the gpt codex CLI? This repository is a tutorial site rather than a package: it holds HTML pages and Markdown, and it has published no GitHub releases, so there is nothing to install from it. Its command line and desktop coverage is instructional, and the third party route it documents for accounts runs through a separate configuration tool whose own releases page it links. ### How do I set up gpt codex? Three paths are given: an official subscription that asks you to supply your own network workaround, a third party relay that the page states is not an official OpenAI subscription, and an install of a separate configuration tool followed by a relay API setup that the table claims carries 150 dollars of daily quota. Fast mode is described as about 1.5 times faster at about 2.5 times the token cost, and the Codex context window is given as 400K tokens. ## Sources - [Issues](https://github.com/xianyu110/gpt-codex/issues) - [Project website](https://codex.maynorai.top/) - [README](https://github.com/xianyu110/gpt-codex/blob/master/README.md) - [xianyu110/gpt-codex on GitHub](https://github.com/xianyu110/gpt-codex) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/xianyu110-gpt-codex