I spend $600 a month on AI programming tools. Two Claude Code (CC) subscriptions for $400, and one Codex Pro for $200. Starting last month, that $200 did more work than the $400.
For the same automation task, CC used $155 of quota, while Codex used $15. Running 7 Codex instances in parallel finished two to three days' worth of work in 25 minutes. After a month of testing: except for web scraping and reverse engineering, Codex is a complete replacement for CC.
I still keep CC mainly out of habit—I've used it for over a year, my entire workflow is built on it, and the migration cost isn't low. But for pure coding tasks, Codex is faster, cheaper, and better. This post details all the pitfalls I encountered and the workflows I developed over the month.
What exactly is Codex?
A programming tool made by OpenAI that directly competes with Claude Code. It also runs in the terminal, can read your code, modify files, and execute commands—its basic capabilities are the same as CC.
But there are a few differences:
Open Source. The code is entirely public, with 75,000 stars on GitHub. CC is not open source. Open source means the community is very active; there are over 50 ready-made skills contributed by others that you can install directly.
Safety Guardrails. Codex's safety guardrails are incredibly strict—a total "moral paragon." It will flatly refuse to help with various scraping or reverse engineering tasks, with cloud-side safety reviews. Currently, bypass communities like linux.do and the Tavern community can't do anything about GPT. Claude Code is relatively more "evil" and will help you do "bad things" 😈.
GPT-5.5. On April 24th, the default model for Codex switched to GPT-5.5, which is a big deal. It's very fast; it starts running seconds after you send a command, reaching about 90 tokens per second even without acceleration mode. The most obvious feeling is: it finally speaks like a human. Previously, GPT-series code comments always had an "AI flavor," but 5.5's output is much more natural.
Which one is stronger?
First, the feel: for daily coding and automation tasks, Codex is significantly faster than CC and saves money—the same task costs $15 on Codex versus $155 on CC. The reason is that GPT-5.5 uses far fewer resources per task than Claude; it accomplishes the same thing with less "thinking."
However, CC has its strengths. For complex refactoring that requires modifying a dozen files simultaneously and understanding the entire project architecture, CC's comprehension is still deeper. In blind tests where people couldn't tell who wrote the code, CC's code quality win rate was 67%. Plus, CC's safety guardrails are weaker, making reverse engineering and scraping a breeze.
Honestly, though, this gap isn't that important in actual use. Later, I'll introduce a tool called Superpowers; once installed, Codex's code quality stability improves greatly, basically closing the gap.
How to Install
I wrote about the installation details in the previous CC intro post, but it's okay if you haven't seen it. Just send "I'm using Mac/Windows, help me install Codex CLI" to any AI (including but not limited to Claude, ChatGPT, DeepSeek, Doubao, Gemini, etc.), and it will teach you step-by-step. If you hit an error, send a screenshot, and it can help solve it.
Codex is even simpler than CC: after installing, type codex in the terminal and log in with your ChatGPT account. No need to configure keys or environment variables; just log in and use.
How Much Does It Cost?
Codex isn't charged separately; it's included in the ChatGPT subscription. The $20 Plus plan works.
However, to be honest, the Plus quota burns very quickly; heavy use can bottom it out in a day. The $200 Pro plan has 5x the usage, and during the current promotion period, it's 10x until June 1st.
My own spending: two CC subscriptions ($400) + one Codex Pro ($200) = $600 per month. Codex's $200 actually gets much more work done than CC's $200 because it consumes 3 to 4 times fewer resources for the same task.
Another key point is API proxies (middleman stations). Codex proxy prices are often 10% of the official price. In my tests, some stations even offer credits at a few cents per dollar. During the Lunar New Year, I even bought a monthly package for 79 RMB that allowed hundreds of dollars of usage daily, though that's gone now. Moreover, Codex proxies are much more stable than Claude Code ones; OpenAI doesn't "glitch out" as often as Anthropic.
However, another point about proxies is that they are semi-underground. The problem with underground industries is that you must test them yourself with real money to find reliable ones. I don't currently have any proxy stations I'm willing to vouch for with my reputation; I've experienced several stations shutting down and have even been scammed out of a thousand or two RMB. I hope everyone tests this for themselves.
First Things to Do After Installation
- AGENTS.md
CC has CLAUDE.md; Codex has AGENTS.md. They function the same—place it in the project root to tell the AI the rules of the project.
Good news: if you already have a CC project, Codex will also read CLAUDE.md and work immediately.
But there's a counter-intuitive discovery: writing too much is actually bad. Research shows that auto-generated AGENTS.md files actually decrease task success rates. This is because Codex can read the code itself; if you stuff it with too many instructions, it gets distracted.
The right way: only write things it can't discover on its own. Build commands, which files not to touch, commit message formats. That's it—no more than one page.
- Permissions: Don't use the defaults
By default, Codex asks for confirmation for every operation, which is annoying. Most developers now use automatic mode—letting it do its thing without constantly pressing Enter.
Set it once in ~/.codex/config.toml:
approval_policy = "never"
sandbox_mode = "workspace-write"
This way, Codex can freely modify files and run commands in the project directory without asking.
But the prerequisite is having Git as a safety net. This is exactly the same as what I said in the CC intro:
Code must be on Git and pushed to the cloud. GitHub, GitLab, whatever. With auto-mode on, Codex becomes more aggressive; I've seen it break config files or delete things it shouldn't. With Git, you can roll back; without it, you have to rewrite.
Databases can't be Gitted; they must be backed up separately. My practice is to back up critical data files every four hours to the cloud. AI sometimes writes scripts that overwrite data—I once had it write a new processing script that directly overwrote the output of a running old script, and the data was gone. Since then, I always make a .bak backup before data file operations.
Summary: Code on Git + push to cloud, scheduled database backups, and .bak before data operations. These three are prerequisites for auto-mode, not options.
How to Assign Tasks
- Let it propose a plan before acting
For complex tasks, don't just let it start changing code. Press Shift+Tab to enter planning mode, let it look at the project and propose a plan, and only let it start once you've reviewed it. This is the same logic as CC—the bigger the task, the more you need to think it through first.
- Be clear about four things in instructions
You don't need to write a lot, but clarify these four: desired result, reference file, what not to touch, and what counts as completion.
Example: "Add rate limiting to the user module. Refer to the auth file's implementation. Don't change existing tests. It's done when all tests pass."
- Useful commands
You'll find that Codex's response quality drops after running for a while. Type /compact to compress its memory or /clear to start fresh. There's also a feature CC lacks: /fork. When unsure which path to take, fork one to try an alternative without affecting current progress.
Worthwhile Tools and Skills
- Money-saving hack: caveman mode
One command can save 65% of quota consumption:
export CODEX_RESPONSE_STYLE=caveman
Once enabled, Codex stops talking nonsense—no explanations, no pleasantries, just work. I now have all my Codex instances set to this by default.
- Superpowers: Quality Stabilizer
I highly recommend this. Simply put, it's a set of rules that forces Codex to work according to a process: think through what to do, write acceptance criteria, write code, and finally check it.
When running "naked," Codex tends to jump straight into code and lose direction. With Superpowers, it's held back by the process; every step has a checkpoint that can't be skipped. The result is much more stable quality for complex tasks.
Installation is simple: download from GitHub and put it in the skills directory. It's not tool-specific—CC, Cursor, and Gemini can all use it because it's essentially just a document.
- Usage Monitoring
pip install ccusage. Once installed, you can see how much quota you burn daily. Without this, you have no idea how much you're spending. I once had a task get stuck in a loop and burn tens of thousands of tokens in minutes; I only realized it because of the monitor.
My Real Workflow
Small tasks: Open Codex directly
Fixing bugs, adjusting formatting, adding tests—just open one codex in the terminal. The simpler the instruction, the better it performs: "Fix the null pointer in this file" or "Add a test for this function."
Big tasks: Split into multiple parallel paths
This is the most powerful way to play. Split a large task into several chunks, assign a Codex to each, and run them simultaneously. The correct way is to create an independent project copy for each Codex so they don't clash.
Actual data: 7 Codex instances working together finished in 25 minutes. A single person working serially would conservatively take two to three days.
The only iron rule: One file can only be assigned to one Codex. Two Codex instances modifying the same file simultaneously will inevitably cause problems, no exceptions.
Pitfalls
- Talking nonsense after a while
Each Codex session has a capacity limit. When it's full, it won't error out; it will start writing unreliable code. This is worse than an error—at least with an error, you know something's wrong. When quality quietly slides, you might have already built several layers on top of its bad code.
My rule: If it starts feeling slightly off, just start a new one. Don't try to extend its life. Better to provide context again than to struggle in a fatigued session.
- Don't install too many extensions
Codex supports external tools, but each tool consumes extra resources. I've seen someone connect a GitHub toolkit with 93 functions, burning over 50,000 extra tokens per dialogue round. Keep only what you truly need and cut the rest.
Real Feelings
Codex is now a complete replacement for CC. I only keep CC out of habit and for scraping needs; everything else is on Codex.
The AI computing power an individual can deploy is increasing rapidly. Last year it was one CC working; this year it's one CC plus seven or eight Codex instances in parallel. If you're using CC and feel it's not enough—add a Codex, and you'll understand what I'm saying.
