OpenAI restores the 5-hour Codex cap: plan for the ceiling

Starting today, the 5-hour rolling usage limit comes back to Codex and ChatGPT Work for ChatGPT Plus subscribers. OpenAI turned that limit off on July 12 and ran about six weeks with only a weekly cap in place. That experiment is over. If your coding agent sits on a Plus seat, your throughput ceiling just moved, and the OpenAI Codex cap is now two numbers instead of one.

What actually happened

Thibault Sottiaux, OpenAI's engineering lead for Codex and ChatGPT, announced the change on X and 9to5Mac reported it on August 24. Plus users now live under a rolling 5-hour window and a weekly allowance. Both apply. Pro tiers stay exempt for the coming months. Regular ChatGPT chat is untouched — this is scoped to Codex and ChatGPT Work.

The stated reason is capacity, phrased two ways. First: the 5-hour window smooths compute load, which is what lets the weekly number stay generous. Second: Plus skews toward newer, more casual users who burn a week of allowance in an afternoon and then hit a wall they didn't see coming. A short window is a guardrail against that.

Both reasons are honest, and both point the same direction. Inference capacity is the constraint, and the pricing tiers are the dial. A limit that came off in July and went back on in August is not a policy — it's a load-balancing lever, and it will move again.

Why metered agents matter for your business

If a Codex seat is doing real work in your shop, you now have a throughput ceiling with a shape: burst-limited by the hour, budget-limited by the week. Those fail differently. A weekly cap fails on the wrong Thursday. A 5-hour cap fails in the middle of a long refactor, and a long-running agent that stalls at hour four is worse than one that never started.

So plan for the ceiling instead of discovering it. Batch the heavy agent work — migrations, sweeps, big refactors — into deliberate windows rather than letting them run whenever someone hits enter. Keep a second route: the Codex CLI on API credits bills per token with no 5-hour window, which is the right home for anything that must finish. And know your actual burn before you upgrade a seat, because Pro's exemption is stated as temporary, not permanent.

The bigger read: any agent that only runs on one vendor's consumer subscription is a workflow you don't control the capacity of. Keep the harness — your prompts, tools, and permissions — separate from the seat that pays for it. Then a limit change is a config change, not a rebuild.

Sources: Thibault Sottiaux on X, 9to5Mac.