Measure your Codex allowance in API dollars from your own session files: weekly burn, reset pace, and whether Plus, Pro or the API is cheapest.
OpenAI shows Codex usage as a percentage bar and a reset date. It doesn't tell you how many tokens a Plus or Pro plan buys, how fast you're burning through the week, or whether paying for the API would be cheaper. This tracker answers those questions from the session files Codex already writes on your computer. It runs entirely in your browser: nothing is uploaded, and only token counts and usage-bar readings are read, never your prompts or code.
Every Codex turn writes a line to a session file under ~/.codex/sessions (on Windows, %USERPROFILE%\.codex\sessions). Each line records the turn's tokens and, usually, the state of the usage bar: used_percent of the weekly window, its reset time, and your plan_type. The tracker finds the moment the bar first reaches each whole percentage, adds up the API-priced cost of the turns in between, and divides by the points crossed. If those per-step amounts are steady, the window is clean and sets your rate. Only clean windows from the last three weeks are used, because the rate shifts when OpenAI changes model generations.
The usage bar counts everything on your ChatGPT account: every computer, every user account on those computers, and Codex cloud tasks. Each computer only keeps session files for its own runs, so a single machine's files undercount. Add each computer's sessions folder one after another, or package it as a .zip or .tar.gz (for example tar -C ~/.codex -czf codex-laptop.tgz sessions) and add the archive. Everything is merged and duplicates are skipped. Windows marked "usage not in these files" mean a step cost almost nothing in your files, which is the sign of a missing machine.
Measured on the author's own account across three computers, each 1% of the weekly bar was worth about $6.40 to $6.70 of API-priced usage on ChatGPT Pro $200 and about $3.30 to $3.80 on Pro $100, with GPT-6 Sol and Astra. That is roughly $650 a week on Pro $200, about 14 times the subscription price at API rates, and break-even against the API lands at about 7% of the allowance. Under GPT-5.6, the same Pro $100 bar was worth about $5.80 per 1%, so re-measure after model changes. Full method and caveats are in the companion article, How much Codex does a ChatGPT plan buy? I measured it.
Plan multipliers (5x for Pro $100, 10x for Pro $200, 20x for grandfathered Pro $200 until Oct 29, 2026, and a reported 25x for Pro $500) come from third-party reports of OpenAI's pricing pages. The weight of GPT-6 Luna on the bar hasn't been measured yet. API prices are list prices and can be edited in the tool. Unused allowance is worth nothing, so a subscription only beats the API once you use more than about 7% of it.
On macOS and Linux they are in ~/.codex/sessions; on Windows in %USERPROFILE%.codex\sessions. If you set CODEX_HOME, they are in $CODEX_HOME/sessions. Choose the sessions folder itself, or drop it onto the tracker.
No. The tracker reads the files in your browser tab and never sends them anywhere. It only reads token counts and usage-bar readings and skips your prompts and code. Refreshing the page clears everything.
Yes. Add each computer's sessions folder one after another, or a .zip or .tar.gz of it, and the tracker merges them and skips duplicates. You need every computer: the usage bar counts all of them, so a missing machine makes your allowance look smaller than it is.
On macOS or Linux run: tar -C ~/.codex -czf codex-laptop.tgz sessions. On Windows PowerShell run: Compress-Archive $env:USERPROFILE.codex\sessions codex-pc.zip. For a server's root account: sudo tar -C /root/.codex -czf codex-root.tgz sessions. Copy the file to the computer you are using and add it.
It is how much usage, priced at OpenAI API list rates, fills 1% of your weekly Codex bar. Multiply by 100 for the weekly allowance. On the author's account it was about $6.55 per 1% on Pro $200 and about $3.50 on Pro $100.
A weekly window where every 1% step of the bar cost about the same amount, with no steps that cost almost nothing. Only clean windows from the last three weeks set your rate, because the rate changed between model generations.
Some 1% steps cost almost nothing in the files you loaded, which means Codex was being used somewhere else at the time: another computer, another user account, or Codex cloud tasks. Add those session files to fix it.
For heavy use, yes: each plan buys roughly 14 times its price in API-priced usage when fully used. Below about 7% of the allowance, pay-as-you-go API is cheaper. The tracker shows where your own usage falls.