The trick everyone's talking about (and nobody explains)
Here's the mistake almost everyone makes. They pick one AI and make it do the entire job. Think of the plan, then write all the code, then fix it, all in one model.
That's like paying a $300-an-hour architect to also swing the hammer, carry the lumber, and paint the walls.
Planning and building are two different jobs. Planning is short and smart. It needs the best brain you've got. Building is long and grindy. It needs a fast, cheap set of hands that won't complain about writing 400 lines of boilerplate.
The move is simple. Use Claude's smartest model to think. Use Codex to do the labor. Each one does what it's actually good at.
Why this actually saves you money
Every AI charges you by the token. Planning barely uses any. A plan is a few paragraphs of thinking.
Execution is where the meter explodes. Reading your files, writing the code, running it, hitting an error, fixing it, running it again. That back-and-forth is 90% of your token bill.
Codex is built for exactly that part. It burns far fewer tokens to get the same job done. So when you hand the execution to Codex instead of making Claude grind through it, the expensive part of the bill gets a lot cheaper.
In my testing this drops overall token usage by about 50% while keeping the output at around 95% of the quality. Same finished project. Half the tokens. That's not a rounding error, that's your $20 plan suddenly doing what the $100 plan used to.
And you never leave Claude. It stays the driver. It plans, quietly hands the building to Codex in the background, then brings the finished work right back into your session.
What you need before you start
Four things, and most of you already have three of them.
- →Claude Code or the Claude Desktop app. This is where everything runs.
- →A ChatGPT account. Even the free tier works. Codex usage counts toward your ChatGPT limits, so a paid plan gives you more room. An OpenAI API key works too if you'd rather pay per use.
- →Node.js 18.18 or later. If you've installed anything technical before, you probably already have it.
- →The Codex CLI. Don't install it by hand. The plugin checks for you and offers to install it if it's missing.
How to install it (the exact steps from my reel)
First, copy this URL. This is the official OpenAI Codex plugin repo:
If you're in the Claude Desktop app (the exact path from the reel):
- Hit the + button
- Go to Plugins, then Manage plugins
- Click Add, then Add marketplace, then Add from a repo
- Paste in the URL
- Click Use
- Click Sync
That's it. Codex can now execute your plans.
If you're in the Claude Code terminal instead, it's three lines:
That last one, /codex:setup, checks whether Codex is ready and offers to install it for you if it's not.
How to actually use it
This is the part that makes it click. Three steps.
- Open a new session and make sure you're using Fable 5 (your smartest planner).
- Paste in your normal prompt for whatever you're building.
- At the very end, add this line:
That's the whole trick. Fable 5 reads your project and writes the plan. Then that one line tells it to hand the actual building over to Codex running GPT-5.5. Codex does the grunt work, Claude brings it back, and you're good to go.
One line. Half the tokens.
When to use it (and when to skip)
This shines on the big stuff:
- →Large, multi-file projects. More building means more savings.
- →Long execution runs. Refactors, boilerplate, anything with a lot of writing and re-running.
- →Any time your token meter is climbing and you start rationing your prompts.
Skip it when:
- →It's a tiny edit or a quick question. The handoff isn't worth it for a two-line change.
- →You want Claude's taste on the final output. For design-heavy or writing-heavy work, let Claude cook.
- →You never hit limits anyway. On a big plan and never run out? This is a nice-to-have, not a must.
Bottom line: if you're building something real and watching your usage, this is free money. Set it up once, add one line, and stop overpaying to build.