How ChatGPT's Limits Really Work in 2026 (Plus, Pro, Free)

Since 2026-08-06, text chats are uncapped on every ChatGPT plan - OpenAI removed the message limits entirely and retired the manually-selected Thinking mode. Free and Go now run on GPT-5.6 Luna, Plus and Pro on GPT-5.6 Sol. What is still limited: images, file uploads, voice, Deep Research, Agent mode, and - since 2026-08-26 - rolling 5-hour windows on ChatGPT Work and Codex for Plus. Here is the current shape of the limits, and how the old three-cap system worked.

Update, 2026-09-18: OpenAI released GPT-6 Astra on 2026-09-03. Regular text chat is still uncapped and still runs on GPT-5.6 Sol for Plus (Luna on Free and Go). Astra itself is limited: Plus gets it only in ChatGPT Work and Codex, at roughly 5-45 messages per rolling 5-hour window. Pro gets it in Chat as GPT-6 Pro with a weekly allowance (50 messages on the $100 tier, shared with GPT-5.6 Sol Pro; 200 on the $200 tier), plus roughly 25-225 ($100) or 100-900 ($200) Work messages per 5 hours. Free and Go have no Astra access. OpenAI also paused new sign-ups and upgrades to the $200 Pro tier on 2026-09-11, with no end date announced; existing $200 subscribers and the $100 tier are unaffected.

The three caps that matter

Until August 2026, ChatGPT's consumer plans had three distinct limit dimensions: a rolling 3-hour message cap on the flagship model, a separate weekly cap on the Thinking reasoning model, and on Free a silent downgrade to a weaker model once the quota was gone. On 2026-08-06 OpenAI removed message caps from text chats on every plan, retiring all three. The table below documents the old system for reference; the limits that remain live sit on images, file uploads, voice, Deep Research, Agent mode, and Codex/ChatGPT Work sessions.

Here is the shape of each cap, the question it answers, and what trips it.

Pro at $200 a month lifts the visible caps, and is for now exempt from the 5-hour gate. But "unlimited" on Pro is a marketing simplification - OpenAI reserves the right to throttle abuse via a fair-use clause buried in the terms. For one human doing real work, Pro is effectively uncapped. For agents and account-sharing setups, it is not.

The 3-hour rolling window, explained (retired August 2026)

The Plus 3-hour cap is the one most paying users hit first. The current count circulated in 2026 by independent trackers is around 160 messages every 3 hours on the GPT-5.x default model. Several 2026 limit guides place it in the 150-160 range, up from roughly 80 in the prior generation. OpenAI does not confirm a precise number anywhere on the help center; you find the real number by hitting it.

The crucial detail: it is rolling, not clock-based. There is no "your 3-hour window starts at midnight." It starts when you send your first message of a new window. If you sent that message at 9:00 AM, your window closes at 12:00 PM. If you sent it at 10:30 AM, it closes at 1:30 PM. After it closes, you get a fresh budget.

This explains a piece of ChatGPT folklore that confuses new users: the "rolling" trap. Someone runs out at 11:00 AM, assumes the limit will reset at 2:00 PM (3 hours later from when they ran out), and is surprised to still be locked out at 2:00 PM. The reset clock counts forward from your first message of the window, not from when you hit the cap.

What burns through a 3-hour window the fastest:

• Long conversations. Every message you send includes the full prior context. A 30-message thread costs more per turn than a 3-message thread, even if the latest message is identical. ChatGPT does not show you a token count, so this cost is invisible.
• File uploads. Plus also caps file uploads at around 80 per 3-hour window. Each upload is processed into the conversation, which makes every later message in that thread heavier.
• Custom GPTs and Operator. Anything you do through a Custom GPT or Operator still counts against the same 3-hour budget. The popular workflow of "let me ask three different Custom GPTs to compare answers" is much more expensive than it looks.
• Asking for Thinking when you do not need it. Each Thinking-mode reply costs a Thinking-bucket message regardless of how short the prompt was. Use the default model when the question is straightforward.

The Thinking weekly cap (retired August 2026)

The Thinking model is the slower, reasoning-heavy sibling of the default GPT-5.x model. It runs longer per response and is the one to reach for on hard analysis, code reviews, and multi-step problems. In 2026 it gets its own separate cap: around 3,000 Thinking messages per 7-day rolling window on Plus, reported by third-party guides.

Three things make this cap surprising in practice:

• It is its own bucket. The Thinking weekly cap is separate from the 3-hour default-model cap. Users on Plus who avoid the 3-hour cap by spacing out their default-model messages often still trip the Thinking weekly cap because they reach for Thinking on every hard question.
• Each Thinking message costs the same regardless of length. A two-sentence question to Thinking still consumes one of your weekly 3,000. The cost is per-call, not per-token, from the user's perspective.
• The reset is 7 days from your first Thinking message of the window. Not a Sunday rollover. If you blew through 3,000 Thinking messages on a Tuesday afternoon, you regain access the following Tuesday afternoon, not midnight on Sunday.

Anyone using ChatGPT for serious coding or research work in 2026 will run into this cap before the 3-hour one. It is the closest equivalent to Claude's weekly all-models cap, and just as easy to hit by accident if you do not have a counter visible.

The Free tier: from silent downgrade to unlimited text (August 2026)

Since 2026-08-06, Free users get unlimited text chats on GPT-5.6 Luna - the silent downgrade for text conversations is gone. Free and Go users also get a per-message Think button for harder questions. Separate allowances still apply to images, file uploads, and voice.

Under the old system, ChatGPT Free behaved differently from Plus: there was no error toast saying "you have hit your limit." Instead, after roughly 10 flagship-model messages in a rolling 5-hour window, the model silently swapped to GPT-4o mini for the rest of the window. You could keep chatting; the replies just got noticeably weaker. The silence of the downgrade (no toast, no banner) is what made new users complain about ChatGPT "getting dumber after a few messages" - they were never told they were talking to a different model.

Pro at $200 a month and the fair-use trap

Pro is the heavy-user tier. OpenAI markets it as having no caps on advanced features, and for normal individual usage that is true: a Pro subscriber doing real work can hammer the flagship model all day without seeing a rate-limit toast. The Plus 5-hour window does not apply, though the weekly allowance still does.

The catch is the fair-use clause. OpenAI's terms allow the company to throttle accounts that show patterns inconsistent with one human's usage - automated requests at machine speed, shared accounts running multiple sessions, requests from many IP addresses in the same minute. In practice, this means:

• For solo professionals, Pro is functionally unlimited. The use cases that hit fair-use are very different from your daily workflow.
• For agentic workloads, Pro is not a replacement for the OpenAI API. If you are running agents that make many parallel calls, the API's published rate limits and pricing are what you should use.
• For shared accounts, Pro can degrade silently. Two people using the same Pro account from different cities can trigger usage-pattern alarms.

Codex, Operator, and Custom GPTs share your Plus pool

One of the most missed details in 2026: everything OpenAI builds on top of ChatGPT shares the same usage pool. ChatGPT.com, the iOS and Android apps, Custom GPTs, Codex and Operator all draw from the same account-level buckets.

If you use ChatGPT for general chat in the morning, ask a Custom GPT to summarise some PDFs over lunch, and run a few Codex prompts in the afternoon, those count as one unified usage stream against your limits. The mental model "Codex is a side product, so it should not affect my main chat budget" is wrong. They are surfaces over the same account.

This is the same shape as Claude's "all product surfaces share one pool" model. The practical implication is the same too: pacing yourself across the day is more important than rationing within any single surface.

How to find out where you actually stand

This is where ChatGPT differs from Claude in a way that matters daily.

Claude has a public usage page: sign in, open Settings > Usage, see your numbers. ChatGPT does not. There is no UI surface in ChatGPT.com today that says "you have used X of Y of this window" - you only find out when you hit the cap and get the rate-limit toast.

That is the gap tokn.watch exists to fill on the ChatGPT side.

Three options to get visibility:

1. Wait for the rate-limit toast. Free, official, but it is the worst possible signal - it only fires when you have already hit the cap.
2. Count manually. Some heavy users keep a tally of Thinking messages in a notebook. Reliable but high-friction.
3. A menu-bar meter. tokn.watch reads your existing ChatGPT.com session cookie, polls the internal endpoints that surface your usage, and shows your remaining percentages plus reset countdowns for whatever windows your account reports, right in your Mac menu bar. Free, no servers, runs alongside Claude tracking if you use both providers.

What to do when you hit a ChatGPT limit

Since 2026-08-06 this section applies to the limits that remain - images, files, Deep Research, Agent mode, and Codex/ChatGPT Work sessions - because text chat no longer has a cap to hit. The action set depends on which allowance you tripped.

Under the old system, if you hit the 3-hour cap on Plus you were looking at minutes-to-hours, not days. The same moves still apply to the remaining rolling windows:

• Wait it out. The window rolls forward. By the time you have made coffee, you usually have a few messages back.
• Switch surfaces, not models. Switching from ChatGPT.com to the iOS app does not give you a fresh budget - it is the same pool. Switching to a different provider (Claude, Perplexity, Gemini) does.
• Drop to GPT-4o or 4o mini for non-critical work. The older models have their own (much higher) limits; you can keep working while the flagship-model window resets.

If you hit the weekly allowance, you are looking at days. The action set is different:

• Ordinary text chat is uncapped, so carry on there for the rest of the week. You can still ask hard questions, you just lose the chain-of-thought reasoning step.
• Switch provider. Heavy users often keep a Claude or Anthropic API account on standby for exactly this.
• Upgrade to Pro at $200 a month. The 5-hour gate is lifted for now. The fair-use clause replaces them, but it almost never trips for solo users.

If you are a Free user noticing the model "got worse", you have been no longer downgraded mid-window: since 2026-08-06 text chat is uncapped on every plan.

Frequently asked questions