The Codex Weekly Limit: How It Works and How to Live With It

Some ChatGPT plans meter Codex twice: a rolling window of roughly five hours sets your pace, and a weekly cap sets your ceiling over a seven-day cycle. The weekly limit is the one heavy users actually plan around, because it is the one that can end your week on Monday. Its size varies by plan and model, OpenAI adjusts it over time, and the current figures live at developers.openai.com/codex/pricing. This page covers the mechanics and the coping strategies; the full limit system is mapped in Codex usage limits, explained.

How the weekly cap works

The weekly limit is a second meter running on top of the short window. Every piece of Codex work, in the CLI, the IDE extension, or cloud tasks, draws from both at once. The 5-hour window refills quickly and bounds bursts. The weekly cap refills slowly and bounds totals.

That layering explains the experience people describe on forums: you can be fine against the short window all day and still get cut off, because the slow meter ran out underneath you. Once the weekly cap is spent, short-window resets stop mattering until the seven-day cycle comes back around.

When it resets

The weekly component resets on a seven-day cycle, not on a fixed calendar weekday. Do not plan around “Mondays”; plan around the timestamp in /status, which shows usage and reset times for both meters on your account. How to read it, plus the other monitoring surfaces, is in how to check your Codex usage.

Why heavy users hit it on day one

Codex meters work, not prompts. Reasoning time, tool calls, and retries all count, and agent workloads multiply each of them. Run three parallel sessions with retries enabled and you are spending at several times the rate a single interactive session would, for hours, unattended.

A hypothetical makes the shape clear. Suppose your plan’s weekly cap is 100 units (an invented number for arithmetic, not a quoted limit). Three parallel agent sessions burning 15 units each per day spend 45 units daily: you are locked out midway through day three, with four days left on the clock. Push the same fleet harder and day one is plausible. Community reports back this up: users describe burning a weekly limit in a single day of heavy agent work, then sitting out the rest of the week.

Usage pattern	What the week feels like
Interactive coding, one session	The 5-hour window pinches occasionally; the weekly cap stays invisible
Daily sequential agent runs	The weekly cap becomes the real constraint by midweek
Parallel agent fleets with retries	Community-reported: cap gone in a day or two, then days of waiting

The weekly limit rewards boring pacing and punishes enthusiasm.

Living with it: four strategies

Watch it. Check /status before you launch anything large, and again the morning after a heavy day. Most lockout stories start with nobody looking at the meter.

Pace it. Defer bulk and low-priority jobs until after your reset timestamp, and let the short window, not the weekly cap, be the thing you bump into. Lower reasoning effort for routine tasks; the deep setting spends the same cap faster for work that did not need it.

Add a lane. A second ChatGPT account you own has its own independent weekly cycle. Two accounts, two ceilings. The accounts must each be yours, with their own subscriptions; OpenAI’s terms prohibit sharing an account between people, and OpenAI keeps the final call over its services. The strategies and the rules are laid out in Codex with multiple ChatGPT accounts.

Automate the switch. Manual lane-juggling fails at 2 a.m. when a cron job hits the cap. Our gateway orders the lanes for you: subscription account A, then account B, then your own OpenAI API key, with the request log naming which lane served each call. That design is covered in fallback lanes, explained.

The honest framing

The weekly cap is not a defect; it is what makes flat pricing survivable for OpenAI while agents multiply usage per person. The practical posture is to treat it like bandwidth: size your plan to your sustained load, keep an overflow lane for bursts, and check the meter before big runs. If you want to know what your current workload maps to in plan terms, the calculator does that math against your actual bill.