ChatGPT's rolling message window, explained

Until August 2026, ChatGPT Plus ran a 3-hour rolling window of roughly 160 messages, starting from your first message of each cycle. On 2026-08-06 OpenAI removed message caps from text chats on every plan, so the window no longer applies to regular conversations. Here is what changed, what is still limited, and how the rolling math worked.

The short answer

What a rolling window actually is

A rolling window is the shape every remaining ChatGPT allowance uses. On Plus the one you meet in practice is the 5-hour Codex and ChatGPT Work window, restored in August 2026; alongside it runs an independent weekly allowance. Per OpenAI's help center article on GPT-5.x in ChatGPT, the plan is designed for high-volume use, and a rolling window is the mechanism that keeps access fair across all users.

The allowance is a request count inside a rolling time window, not simply a hard time limit. Think of it as a budget that renews every 5 hours: you can spend it all at once, or stretch it across the full window - either way, it resets 5 hours after your first request of the current cycle.

The key point that surprises most new Plus users: there is no built-in usage page in ChatGPT. Unlike Claude, which shows your session percentage and reset time at Settings > Usage, ChatGPT gives you no visibility into how many messages remain in your current window. You find out you have hit the cap by hitting it. This is the primary reason Plus users look for external trackers - the information simply is not surfaced before the wall.

The rolling-window math

The rolling anchor is what most users get wrong. The 5-hour clock starts from your first request of the cycle - not from when you exhausted the budget.

The rolling trap catches people when they hit the cap mid-morning. Someone thinks: "I ran out at 12:30 AM, so my limit resets at 2:30 PM - three hours from now." But the reset clock started at 10:00 AM when they sent their first message. The window closes at 1:00 PM - 1.5 hours sooner than expected. Misreading the anchor is the single most common reason Plus users wait longer than they need to.

ChatGPT does give you the remaining wait time in the error toast - something like "You can send more messages in 1 hour, 22 minutes." That relative time counts down from the first-message anchor. Trust that number, not your own arithmetic from when you hit the cap.

The earlier you exhaust the budget, the longer the wait. The later you exhaust it, the shorter the remaining window - because the 5-hour clock has been running the whole time, regardless of whether you were using ChatGPT or not.

What counts against the 5-hour window

The Plus 5-hour window covers Codex and ChatGPT Work usage across your account. These all draw from the same pool:

• Messages on chatgpt.com - every message you send and receive on the web interface.
• Mobile apps - iOS and Android ChatGPT apps share the same pool as the web app.
• ChatGPT macOS desktop app - separate app, same account, same pool.
• Every surface - web, the mobile apps and the macOS desktop app all read the same per-account budget. Switching between them does not give you a fresh one.
• Codex sessions tied to your ChatGPT account - agentic coding tasks that run multiple steps burn through the budget at the same rate as regular messages. A Codex run that fires 10 tool steps looks like 10 messages to the cap.
• Operator - automated sequences triggered via Operator draw from the same pool.
• File uploads - uploads have their own allowance, separate from the 5-hour window. Each upload processed into a conversation also makes later requests in that thread heavier, so uploads compound the drain on whatever budget you are spending.

What does not count against it

Two common points of confusion about what shares the pool:

• Ordinary text chat does not count against it. Since 2026-08-06 text chat has no cap at all, on any plan. You can exhaust the 5-hour Codex window and carry on chatting normally.
• Images, uploads, voice and Deep Research have their own allowances. Each is independent of the 5-hour window and of the others.

This separation is what makes the limits survivable. Exhausting the 5-hour Codex window costs you agentic coding for a few hours, not the ability to use ChatGPT. The full mechanics of all three limit layers are covered in how ChatGPT's limits really work in 2026.

What does not reset the window

If you have hit the cap and are waiting for the reset, none of these shortcuts work:

• Logging out and back in. The cap is enforced server-side at the account level, not in the session.
• Starting a new chat. New conversations do not get a fresh message budget. You have one pool per account, not one per chat.
• Switching from chatgpt.com to the mobile app or to the macOS desktop app. They all read from the same per-account state.
• Switching from a Custom GPT to the main interface. Same pool.
• Refreshing the page.
• Waiting until midnight. The window is rolling and anchored on your first message. No calendar boundary resets it.

The only mechanism that resets the window is the 5-hour clock completing from the time of your first request. The error toast tells you the remaining wait; that number is the ground truth.

How firm are the published numbers?

The published ranges for GPT-6 Astra in Codex and ChatGPT Work are roughly 5-45 requests per 5 hours on Plus and Business Standard, 25-225 on Pro at $100 and 100-900 on Pro at $200, with Business Premium exempt from the 5-hour gate. The ranges are wide on purpose: a request's cost depends on the work it does, so a long agentic run spends far more than a short one.

OpenAI has adjusted these numbers without notice in the past, both upward when capacity expanded and downward during heavy usage periods. Treat the top of a range as an optimistic ceiling for planning, not a guarantee. If you run out sooner than the range suggests, your requests are simply costing more than the cheapest ones.

"Limits for Plus are designed to provide a great experience for all ChatGPT users. We periodically adjust these limits based on usage and capacity." OpenAI, ChatGPT Plus documentation (paraphrased)

The 5-hour window vs the weekly allowance: how they interact

Most Plus users who do analytical or coding work will eventually hit both caps. Understanding the interaction prevents confusion when one resets and the other does not.

The weekly allowance is the one heavy users hit without expecting it, because it accumulates quietly across the week while the 5-hour window keeps resetting in front of them. When the weekly allowance runs out, waiting for the 5-hour window to roll over does not help: the weekly one is what is stopping you, and it does not lift until its own seven days complete.

The transparency gap: ChatGPT vs Claude

The most significant practical difference between ChatGPT and Claude for users who care about caps is visibility. Claude gives you a Usage page at Settings > Usage that shows each cap's percentage consumed and the exact reset timestamp - you can check it any time before hitting the wall. Claude Code has the /usage command for the same information in the terminal.

ChatGPT has no equivalent page. There is no percentage display, no countdown timer, no "X messages remaining" anywhere in the interface while a window is active. The only moment OpenAI communicates the cap is when you hit it - the error toast appears and shows the remaining wait time as a relative countdown. Once you dismiss that toast, the information is gone until you hit the cap again.

The practical workaround most Plus users land on is manual counting. The more structured workaround - and the one that scales to heavy use - is a passive menu bar monitor that reads the session state and surfaces the window continuously in your peripheral vision.

Frequently asked questions