OpenAI Codex Tutorial: From Installation to Practice — Nine Core Features Explained

OpenAI Codex has been on a roll lately, with frequent feature updates and Sam Altman raising users' weekly usage limits twice in a row. Compared to Anthropic's Claude Code, which requires command-line operation, Codex offers a full visual desktop application that's very friendly even to non-programmers. Based on hands-on experience, this article systematically covers Codex's core features and usage tips across nine modules to help you quickly master this AI coding tool.

Background: OpenAI Codex was originally released in 2021 as a code generation model and served as the underlying engine for GitHub Copilot, trained specifically on billions of lines of public code based on the GPT architecture. The new generation of Codex in 2025 has evolved from a simple code completion model into a full AI programming Agent platform, integrating task planning, tool invocation, long-term memory, and other capabilities. It represents a paradigm shift from "code completion" to "autonomous development" — it can not only "say" what to do but actually "do" it autonomously.

Installation and Login

The installation process for Codex is very straightforward. Visit the Codex official website, choose the Windows or macOS version based on your operating system, and download and install it. After installation, a login prompt will appear on your desktop — just sign in with your OpenAI account to get started.

Note that free users have very limited usage and may run out after just a few messages. To properly use Codex's full features, subscribing to a ChatGPT membership is recommended. With the Pro plan, for example, after using 400 million Tokens in a week, 82% of the quota still remained — quite generous. This is partly thanks to Altman raising the limit cap twice in a row.

What are Tokens?: Tokens are the basic units that large language models use to process text, and they don't strictly correspond to words or characters. In English, one Token equals roughly 0.75 words; in Chinese, one character typically corresponds to 1-2 Tokens. A weekly usage of 400 million Tokens is enormous — assuming an average of 2,000 Tokens per conversation, that's approximately 200,000 conversations. More powerful models like GPT-5.5 typically consume more Tokens because their reasoning chains are longer and context windows are larger, which is why the fast mode consumes more quota.

Basic Interface and Chat Area

Codex's interface layout is similar to the ChatGPT web version: the left side has a task bar for managing various tasks; the center is the main chat area; and the right side is a multi-function panel for the project where you can view files, open a terminal, start sidebar chats, and more.

In the chat area, there are several key settings worth noting:

• Model Selection: Ranging from GPT-5.5 to GPT-5.2, covering various main models and branch models. Intelligence levels from low to high represent reasoning intensity.
• Speed Mode: Standard speed and fast mode. Fast mode consumes more quota.
• Permission Management: Three permission levels — default permissions (sandbox execution), auto-review (sandbox + automatic review), and full permissions (direct computer control). Choose based on your needs.

Sandbox Mechanism Explained: A sandbox is a security isolation mechanism that runs programs in a restricted virtual environment, preventing them from accessing or modifying critical resources on the host system. Codex's default permission mode runs in a sandbox where AI-generated code can only execute within the isolated environment and cannot directly read/write the user's file system or access sensitive network resources. This design strikes a balance between security and functionality — for uncertain code tasks, sandbox mode is the safest choice. Only when you need Codex to truly control your computer should you enable full permissions mode.

Typing a forward slash / in the chat box brings up various function commands, including viewing remaining context length, model limits, refresh times, and other status information.

Multi-Task Parallelism

One of Codex's most practical features is multi-task parallelism. In traditional development, waiting for one task to complete before starting the next is inefficient. In Codex, you can open multiple task branches simultaneously within the same project.

For example: one task handles building a website template while another gathers Codex news from the internet — two isolated tasks running in parallel without interfering with each other. When a task completes early, a small blue dot notification appears in the left task bar. Completed tasks that are no longer needed can be archived directly to keep the workspace clean.

Codex also has a built-in "pet" feature that displays task progress in real time. Whether you're browsing Bilibili or scrolling through TikTok, you can preview project progress alongside and receive notifications when tasks are completed.

Image Generation: Built-in Image Gen 2

Codex has GPT Image Gen 2, a top-tier image generation model, built in. Type /image in the chat box to invoke it.

It might seem like a simple drawing feature, but it's enormously valuable in real projects. When you need design assets, icons, banners, and other materials, you can generate them directly in Codex with fully controllable dimensions and styles, eliminating the need to find a separate designer. This "development + design integrated" workflow offers a very noticeable efficiency boost for indie developers.

Plan Mode: From Verbal Description to Executable Plan

Plan mode is an extremely important feature in Codex's coding workflow, and Claude Code has a similar mode. Its core value lies in: transforming your verbally described requirements into a complete development plan that the Agent can accurately understand and execute.

The Difference Between AI Agents and Traditional AI Assistants: The core difference between AI Agents and traditional conversational AI lies in "autonomous action capability." Traditional AI assistants can only generate text responses and require humans to manually execute suggestions. AI Agents, on the other hand, can autonomously invoke tools, execute multi-step tasks, perceive environmental feedback, and adapt their behavior accordingly.