OpenAI Brings Back the 5-Hour Codex Cap for Plus and Business

If you spent the first week of September sailing through Codex on ChatGPT Plus, expect the session meter to tighten again. OpenAI has reimposed the 5-hour rolling session cap on Codex and ChatGPT Work for Plus and Business standard users, ending the looser window that came with the GPT-6 Astra rollout, and it staged a global usage reset for every paid subscription this week so subscribers could ramp straight into the new model on a clean meter.

For anyone carrying a heavy agentic coding bill, this matters more than the timing of a single reset. It tells you the plan tier you actually need, and it changes how you should schedule long Codex runs so a promotion that looked unlimited stops cutting you off mid-context.

The 5-Hour Codex Limit Is Back for Plus and Business

The 5-hour limit is a rolling, short-window gate, not a weekly quota. OpenAI lifted that short-window cap in mid-July around the GPT-5.6 Sol rollout, leaving only the weekly usage limit in place for Plus, Pro, and Business plans, according to a support-side post on the OpenAI Developer Community. For about six weeks, developers got used to sessions that ran as long as their weekly allowance allowed, without a separate meter stopping them for using their quota in one sitting.

That grace period is what OpenAI is now walking back. Thibault Sottiaux, the engineering lead for Codex and ChatGPT Work, announced in late August that the 5-hour limit would return for Plus accounts across Codex and ChatGPT Work. His stated reasoning was compute: the rolling window lets OpenAI smooth demand and still keep Plus generous on its weekly allowance. He was also explicit about the product problem underneath, some Plus users were burning through a whole week of usage in a single overnight sprint and then getting confused about why they had nothing left for the rest of the week.

The Astra launch in early September then gave everyone a few extra days of looser behavior, with usage resets arriving frequently enough that some Power users started treating the reset schedule as free compute. This week OpenAI is closing that window for Plus and Business standard: the 5-hour rolling cap applies again on top of your weekly limit, and a one-off global reset was lined up around 6pm PST on September 7 so paid accounts start the next cycle fresh.

Two Meters, One Shared Pool, Two Different Resets

The part most heavy users underestimate is that a ChatGPT Plus quota is not one bucket. OpenAI’s help center now states that usage from Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents draws from the same agentic usage and credit pool on the plans that enable them. But inside that shared pool you are governed by two separate meters with independent clocks.

One is the weekly limit, the headline allowance you see in your plan. The other is the 5-hour rolling window, a shorter throttle on how fast you can spend that allowance. Hitting the 5-hour meter stops a Codex session even when you still have weekly usage left. And as developers have been quick to point out, the weekly reset does not refresh your 5-hour counter, so you can watch your allocation replenish while your burst window stays closed.

That asymmetry is the real cost driver. If you code in dense weekend spurts, the effective ceiling on your work is not your weekly total, it is how much of that total you can fit inside a single 5-hour window at a time. Someone who drips usage across a whole week gets more done on the same dollar than someone who front-loads, purely because of the time gate, not the token price.

ChatGPT Plus Usage Limits Now Gate Your Agentic Runs

This is precisely where the upgrade pressure shows up. With Astra now available to subscribers and priced like a premium workhorse, the tier above Plus has quietly become a more compelling buy. Developer sentiment after the reimposition sums it up well: the 5-hour session gate is a lazy way to shift the burden of continuous operation back onto the user, and the paid ladder is what you pay to escape it.

For the TokenKarma reader the useful framing is structural, not just defensive. OpenAI is now sending the same signal Anthropic sends when it tightens the burst behavior around Claude Code: the top consumer tiers are treated as gateways to a more expensive plane where your sustained, continuous work is actually allowed to live. Both vendors have realized that a hard ceiling on burst sessions is the cheapest demand-control they have, far cheaper than building out enough idle capacity to serve every subscriber at peak all at once.

What the Reimposed Codex Limits Mean for Heavy Users

If you are on Plus or Business standard and do sustained agentic work, treat the 5-hour window as a hard engineering constraint rather than a soft guideline.

Schedule bursts inside the window. Because the weekly reset does not reset the hourly meter, plan your highest-value runs to start right after your burst window opens rather than whenever the wall clock is convenient. A run that crosses the 5-hour boundary gets cut at the moment your context is deepest.

Watch what you lose at the cut. A session stopped by the rolling meter does not just pause your token spend. If you were holding a long working context with cached state, state that you would have reused in a longer run, an interrupted agent needs to rebuild it, and rebuilding can cost more than the tokens that were already spent. Keep a restartable checkpoint for anything you intend to resume.

Do not budget on the reset. The global reset this week is a launch-week artifact tied to Astra, not a sign that OpenAI is moving to unlimited metered use. Model your long-run spend on the real terms: weekly allowance plus a 5-hour burst ceiling. If your actual workload needs more sustained throughput than that, the honest fix is moving mission-critical continuous coding to a higher paid tier or to an API plan where throughput is per-account and you control concurrency.

Re-check where sessions actually run. Because Codex, ChatGPT Work, and Workspace Agents all draw from one agentic pool, a single chat agent background job can silently consume the same allocation that your foreground Codex session needs. If you run both, treat them as one budget or you will keep getting throttled at the exact moment one of them is doing real work.

The Playbook for the Gate Week

Do not cancel anything on impulse; this is a change in cadence, not a kill switch. Re-baseline your weekly ceiling on the assumption that 5-hour windows, not weekly totals, define how much you can really ship. If you were one of the users who “pharmed” the frequent resets as free capacity, stop betting on them now that the mechanism is back. And if the gate genuinely caps a workflow you rely on, that is the signal to price out the tier that fits, not to pay more for the same throttled meter. The most expensive mistake after a reset-heavy launch week is assuming the good times were the real pricing model. The real model is the ceiling that is now back in place, and it is best to budget for the ceiling from the first day.