Codex Rate Limits, Explained
Two questions come up constantly about Codex limits: why did I get cut off when the chat sidebar said I still had messages left, and why does the number on my phone not match the number in the CLI. Both come down to the same misunderstanding — Codex usage and ChatGPT chat usage are not the same pool.
Two separate limit systems
Sending messages to ChatGPT in a normal conversation draws from a per-model message allowance. Running Codex — in the CLI, in the IDE extension, or through the cloud agent — draws from a completely separate usage budget, measured in two rolling windows: a short one (around five hours) and a longer one (about a week). Hitting zero on one doesn't touch the other. If chat still works but Codex refuses to run a task, or the reverse, that's expected, not a bug. Worth knowing if you also use Claude: Claude Code works the opposite way — it shares its limit with claude.ai chat instead of splitting it.
Why the windows reset differently
The 5-hour window is the one you'll hit first during an active work session — it's what stops a long agent run partway through. The weekly window is the backstop: even with the 5-hour window fully available, you can still be capped for the week. Both show a reset time, and both are visible from the same two places: the CLI's /usage command, and Settings → Usage on chatgpt.com.
The allowance depends on your plan, and it moves
Free, Plus, Pro, Business and Enterprise accounts get different multiples of the base allowance, and OpenAI has adjusted the specific numbers repeatedly over 2026 — pausing the 5-hour limit outright for a period, resetting weekly limits multiple times after fixing consumption bugs, and adding a "banked reset" system that lets you save an unused reset and spend it later instead of losing it. Any specific number written down here would likely be wrong within a few weeks. The reset time and remaining percentage on your own account, checked at chatgpt.com or via /usage, is the only number worth trusting.
What actually helps
Since the caps move, watching a static number less often matters less than watching the rate you're burning through it. That's the one thing a snapshot percentage can't tell you on its own — 40% used could mean "fine" or "gone in ten minutes" depending on how fast it's dropping. CapMeter tracks that rate continuously in the menu bar for both Codex and Claude accounts, which is the reason it exists instead of a page that just mirrors the settings screen.