The Best ChatGPT Codex Setup for Most Professionals

I burned through 25 billion tokens before I fixed this

I checked my Codex dashboard one month and found I had used 25 billion tokens. That is roughly 5 billion a month, with a peak of 3.9 billion in a single week and 798 million in one day.

The number itself was not the problem. The problem was what I was paying to support it, and how much more I could have paid by following the obvious advice.

The obvious advice is usually wrong here. Most people in my position get told to either run the API directly or buy the most expensive plan. Both are expensive mistakes.

The API looks cheap until you run your real volume

The API is priced per million tokens, and at low volume it feels like nothing. At my volume it becomes the most expensive option by far.

Luna costs about $1 per million normal input tokens and $0.10 per million cached. Terra costs $2.50 and $0.25. Output and reasoning tokens cost more on top.

• Best case at 5 billion tokens. If every token were the cheapest cached input, the Luna API would still run about $500 a month. Terra would run about $1,250.
• Peak week pace. At my 3.9 billion token week scaled to a month, the Luna API reaches roughly $1,690 and Terra roughly $4,225.
• Reality is worse. Some tokens are normal input, cache writes, output, or hidden reasoning. The dashboard does not show the split, so the real bill is higher than these floors.

The API is the right tool for a product with a cost model. It is the wrong tool for your own daily coding habit.

None of this counts the mental overhead of watching a meter and rationing work. That cost is real even when the dollar cost is not.

The $100 plan is the one that fits serious usage

Three individual tiers exist, and only one sits at the price that matches heavy Codex work without overshooting it.

• Plus, $20 a month. One times the base Codex allowance. Fine under light use.
• Pro 5x, $100 a month. Five times Plus for an extra $80. This is the workhorse tier.
• Pro 20x, $200 a month. Twenty times Plus. Most people never need it.

The estimated five-hour limits show the gap. On Pro 5x, Luna runs about 250 to 1,400 tasks per window, Terra about 100 to 550, and Sol about 75 to 450. Plus caps those far lower, and Pro 20x pushes them higher than most solo workflows can fill.

If Plus rate limits keep stopping your work, or you are already paying Plus overages past $80 a month, Pro 5x is the move. Save Pro 20x for the rare case where you are on Pro 5x and still spending more than $100 a month on extra credits.

Set the model by the stakes, not by reflex

Buying Pro 5x is half the fix. The other half is telling Codex which model to use and when.

Most of my early waste came from two habits: reaching for the biggest model by default, and leaving Fast Mode on. Both are easy to undo.

• Luna, medium reasoning, by default. Routine edits, UI changes, debugging, and well-scoped features. This is most of the work.
• Terra, medium or high reasoning, when needed. Architecture, ugly bugs, and changes that touch many files at once.
• Sol, only when Terra fails or the result really matters. Keep it in reserve, not in the rotation.
• Standard speed, always. Only switch to fast when the wait is the actual blocker, not the vibe.

Cheap models where cheap mistakes are survivable. Frontier only where being confidently wrong gets expensive.

Max and Ultra are not for ordinary tasks. Leave them off. The point is to keep the frontier model for the moments that justify its cost, not to spend it on a rename across six files.

Fast Mode is the leak you do not see

I had Fast Mode on for 59 percent of my usage. That single setting was likely wasting more allowance than any model choice.

On supported models, Fast Mode spends about two to two and a half times the normal credits for roughly one and a half times the speed. You pay a large premium for a small clock change.

Turn it off. Turn it on only when the speed is the real constraint, finish the urgent thing, and turn it back off. Treated as a default, it is the most expensive habit in the setup.

The rule that keeps you on the right tier

The setup is not a one time decision. It is a small rule you check when the bill or the limits tell you to.

1. Stay on Plus while you spend under $80 a month on extra credits and can live with the occasional reset.
2. Move to Pro 5x when Plus overages hit $80 or rate limits regularly stop your work.
3. Move to Pro 20x only after Pro 5x, and only if you still spend more than $100 a month on extra credits.

For my usage the answer was clear: Pro 5x at $100, Luna by default, Terra when necessary, Sol when Terra fails, and Fast Mode off. Fewer interruptions. No $4,000 API bill. No plan I could not fill.