Codex Beginner's Hands-On Guide: From Installation to Million-Token Context

After OpenAI deeply integrated Codex with ChatGPT, many users are still confused about downloading, account setup, and feature usage. This article is based on a hands-on tutorial from a Bilibili creator, walking through the complete workflow from installation to advanced Codex usage — and analyzing the trade-offs behind several key features.

A heads-up: some account registration and subscription methods mentioned in the tutorial involve third-party channels, which carry compliance and account security risks. This article is for informational purposes only and does not constitute operational advice.

The Product Logic Behind the Codex–ChatGPT Merger

The tutorial repeatedly emphasizes one premise: Codex and ChatGPT have been merged, and the GPT client you download already includes Codex functionality. After opening the client, you'll see two entry points — Codex and ChatGPT — each representing a different task-handling paradigm.

For simple Q&A needs, switching to ChatGPT mode is sufficient; responses are fast, typically arriving within seconds. For scenarios requiring specific task execution and continuous computer operations, use Codex — processing times can range from a few minutes to tens of minutes. This division essentially distinguishes between two interaction paradigms: "conversational AI" and "agent AI." The former is instant question-and-answer; the latter involves handing a goal to the model to autonomously break down and execute.

For beginners, understanding this distinction matters more than memorizing installation steps. Whether you're chatting with an assistant or delegating tasks to an automated agent that can control your computer determines which entry point to use and which permissions to enable.

The concept of "AI Agent" deserves a dedicated explanation. Unlike traditional chatbots, an AI Agent can autonomously plan subtasks, invoke tools (such as a browser, file system, or code execution environment), and dynamically adjust its course of action based on intermediate results — all without requiring step-by-step user intervention. OpenAI positions Codex within this framework as a "Computer Use Agent" that can directly control the mouse, keyboard, and open applications, effectively treating the user's desktop as an execution environment. This is an entirely different product from the earlier Codex code-completion model (released in 2021, discontinued in 2023) — the reuse of the name causes easy confusion. The current Codex is closer in nature to Anthropic Claude's Computer Use feature or Microsoft's Copilot Actions. Its core capability lies in the closed loop of "perceive screen → understand task → generate action → execute and respond," rather than simple code generation.

Tips and Key Settings

When you start a new Codex conversation, the most prominent element on the interface is the "Full Access" button, which determines whether Codex needs user approval before executing actions. The creator's recommendation is to leave it off at first, spend a day getting familiar with the software, and then enable it — at which point you can "hand off a task and not worry about anything." This gradual authorization approach is beginner-friendly: observe the agent's behavior first, then progressively unlock permissions, rather than immediately granting the AI high-level control over your computer.

For model selection, regular subscribers can use mid-to-low tiers for everyday tasks and scale up for complex ones, avoiding rapid credit consumption. Several advanced features worth noting:

• Goal Mode: Sets a target for Codex that it won't give up on until achieved. The creator mentioned a Twitter case where a user was "woken up" by Codex in reverse — to get the person to enter a password, the model wrote a program that made the computer emit sounds to wake them up. This example shows that agent behavior in Goal Mode can be quite aggressive. It's not recommended unless you're certain the goal is achievable and you're comfortable with the cost.
• Plan Mode: Breaks a task into a dozen or so steps to be executed in sequence. For complex tasks, you can manually modify steps to improve execution success rates.
• Record Skills: Lets Codex watch a screen recording and extract a reusable skill set from it. Best suited for fixed, repetitive operations or tasks that can be distilled into a standard SOP; effectiveness is limited when real-world information exchange is involved.
• Branch Feature: Lets you open multiple branches without interfering with the original conversation, executing unrelated tasks simultaneously. In theory, this can boost efficiency by an order of magnitude — at the cost of proportionally higher credit consumption.

In the settings, the creator recommends enabling auto-review permissions, context window usage display, web search, and local memory, as well as installing relevant components under computer control to strengthen Codex's task-handling capabilities.

The "credits" mechanism is an unavoidable cost variable when using Codex. OpenAI uses a compute-consumption billing model for agent-type operations: the higher the model tier, the longer the context, and the more parallel branches — the more credits a single task consumes. Take the branch feature as an example: opening a dozen parallel tasks theoretically multiplies throughput several times over, but at a proportional token cost. For subscription users, this means monthly quotas can be exhausted quickly. A sensible strategy is to tier tasks by urgency: use ChatGPT mode (zero or low cost) for everyday queries, run mid-complexity tasks on a lower-tier model in single-thread mode, and only upgrade tiers or enable branches when parallel processing or long-document handling is genuinely needed. Making a habit of enabling "context window usage display" in settings helps you monitor consumption in real time and avoid quality degradation from automatic compression when limits are exceeded.

Plugins and Million-Token Context

Plugins are essentially skill extensions for specific scenarios. For example, an Excel plugin enhances Codex's ability to handle spreadsheet documents, and users can install them as needed. The creator particularly recommends a "simulated click plugin" that allows Codex to control the computer's mouse more precisely, improving the quality of task outcomes.

More noteworthy is the adjustment of context capacity. Codex's default context capacity is 200K tokens; once exceeded, it automatically compresses, which weakens task execution quality. The tutorial provides a method to expand context to one million tokens: the official default has this disabled, but it can be enabled by modifying a configuration file.

The specific approach is to have Codex open the relevant configuration file directly, locate it in the file manager, paste the instruction and save, then restart. After restarting, you can see the context capacity displayed in the conversation (the tutorial shows 824K), confirming the expansion worked. There's a cost trade-off here as well: enabling million-token context will significantly increase credit consumption, and the creator recommends enabling it selectively.

The context window refers to the maximum number of tokens a model can "see" and process in a single conversation. It directly affects the quality of long-document comprehension, multi-turn task memory, and complex code analysis. When task content exceeds the window limit, the system compresses or truncates earlier content, causing the model to forget critical details and leading to execution errors or drift. Codex's default 200K token limit is sufficient for most everyday tasks, but may fall short when handling large codebases, lengthy documents, or multi-step tasks running over several hours. Enabling million-level context by modifying the config file essentially allows the local client to break through the conservative official default, retaining more conversation history and task state in the active context. Keep in mind that a larger context means the model must process more tokens on each inference pass — computational cost and response latency rise accordingly, making it not worthwhile for every task.

A Starter Project: Build a Desktop Pet

The tutorial concludes with a hands-on project: using Codex to create a desktop pet. The process involves clicking the avatar to enter Codex settings, finding the "Pet" option, clicking to create, stating your requirements in the dialog box that appears, and uploading assets — Codex will then generate a desktop companion.

The creator's finished product features transparency effects and multiple skins, and is connected to DeepSeek^Ad for real-time reading of computer information and conversational interaction. The value of this project is that it ties together all the earlier settings — permissions, plugins, context — into a concrete output, giving beginners a much better sense of Codex's capability boundaries than simply following buttons.

Usage Recommendations and Risk Warnings

Overall, the core takeaway from this tutorial can be summarized as follows: Codex is a capable agent tool, but its powerful features generally come with credit consumption and authorization risks that users must weigh against their own circumstances. The extreme case with Goal Mode, the gradual enabling of Full Access, and the cost implications of million-token context all reflect this logic.

The creator also warns at the end that Codex is powerful, but "don't overindulge" — they personally stayed up too late and harmed their health from overuse. Beyond that caution, it's worth emphasizing further: the tutorial mentions third-party account registration, bypassing phone verification, and topping up memberships through unofficial channels — all of which carry clear compliance and account security risks. Readers should prioritize official channels and carefully evaluate the risks involved.