Codex & ChatGPT Usage Limits Explained

How Codex's usage limits actually work

OpenAI meters Codex CLI, the Codex IDE extension, Codex on the web, and Codex's cloud tasks against your ChatGPT plan through two windows stacked on top of each other. A rolling 5-hour window doesn't reset at a fixed clock time; usage from five hours ago frees up continuously as you go. On top of that sits a weekly cap that resets seven days after your first request in the current period, and the new period only starts once you send another request after that reset lands. Local CLI messages, background cloud tasks, and ChatGPT Work sessions all draw from the same shared pool, so a big job you kicked off in the morning can leave less room for the debugging session you actually wanted the window for that afternoon. Hit either window and Codex stops taking new work until it refills.

Since April 2026, both windows meter tokens and compute rather than message count, so a short "fix this typo" prompt barely registers while a long agent run that reads a dozen files and re-runs the test suite can eat a real chunk of the window in one go. Inside the CLI, /status shows both percentages and when each one resets, and turning on five-hour-limit and weekly-limit in /statusline keeps the numbers visible while you work, which matters because they move constantly.

What changed in 2026 (and why the limits keep moving)

OpenAI turned the 5-hour window off entirely for Codex and ChatGPT Work in mid-July 2026, leaving only the weekly cap in place. It brought the window back for Plus subscribers on August 25, saying the pause was about managing load on its compute and that casual Plus users were burning through a whole week's allowance without realizing it. Pro's $100 and $200 tiers were still running on the weekly cap alone going into September, and OpenAI has described that gap as temporary rather than a permanent split between plans.

The exact dates matter less than the pattern: Codex's limits have shifted more than once in 2026, each time through a help-center update rather than a pricing change, so a number you read last quarter can already be stale. Claude's usage limits have moved in similar ways, which suggests this is just what metered coding-agent subscriptions do once heavy agent use shows up in the usage numbers, not something specific to OpenAI.

Plus vs Pro vs Business: what each plan actually gets

"Not enforced as of Sept 2026" is doing a lot of work in that row. OpenAI has already changed which plans carry the 5-hour window twice this year, so treat this table as a snapshot rather than a promise, and check OpenAI's own Codex plan page before budgeting around it. Business, Enterprise, and Edu accounts run on a separate credit-and-overage system instead of a hard weekly stop, which is a different arrangement worth reading up on separately if that's your plan.

Why agentic coding burns through the window fast

Token-based metering is why the weekly cap can run dry before the 5-hour window ever fills once. A handful of large agentic runs do more damage than a day of short chats, because everything the agent reads and re-reads counts against the window, not just what it writes back. One thread on OpenAI's developer forum is a good example: a user drained the entire weekly cap with the 5-hour window barely touched, all from a small number of heavy sessions.

• Long test-and-fix loops, where the agent reruns the suite after every change
• Large multi-file refactors that re-read the whole directory on each turn
• Long conversations that keep re-sending earlier context instead of starting fresh
• Background or cloud tasks running at the same time as CLI work, since they share the same pool

A second lane for the everyday work

None of this is a case against Codex. For the genuinely hard problems, the plan you already pay for is worth keeping. The friction shows up on the volume behind those problems: writing tests, running a multi-hour refactor, cleaning up docs, letting an agent loop for an hour while you do something else. That's exactly the kind of work that drains a 5-hour window fastest, and exactly the kind of work a second, flat-rate lane is good for.

NeverMetered runs one model, Qwen3.8-27B, on its own dedicated servers, with no token quota, no 5-hour window and no weekly cap. Instead of metering usage, it prices by how many requests can run in parallel at once; anything past that count waits briefly in a fair queue instead of failing. It speaks the OpenAI and Anthropic APIs on one base URL, so pointing Codex CLI at it, or any of the other supported tools, is a config change rather than a rewrite. See current plans for stream counts, or read how NeverMetered runs for the full picture on the model, the servers, and what happens to your prompts.

It won't replace Codex for your hardest debugging session, and it isn't trying to. It's somewhere to send the rest of the work, so the meter you're already paying for stays free for the parts that actually need it. If you're weighing this against sticking with one plan, the how NeverMetered compares hub walks through it side by side with Codex and a few other tools.