The GPT-5 Thinking weekly cap nobody told you about

Until August 2026, ChatGPT Plus gave you roughly 3,000 manually-selected GPT-5 Thinking messages per week, on a rolling window that started from your first Thinking message. On 2026-08-06 OpenAI removed message caps from text chats on every plan and retired the manual Thinking mode in favor of a thinking-effort control. Here is what changed, and how the old cap behaved while it existed.

Update, 2026-09-18: A weekly allowance is back in ChatGPT, but only for one model. GPT-6 Astra (2026-09-03) appears in Chat on Pro plans as GPT-6 Pro with a weekly message allowance: 50 on the $100 tier (shared with GPT-5.6 Sol Pro) and 200 on the $200 tier. Plus has no weekly cap; its Astra access is in Work and Codex only, at roughly 5-45 messages per rolling 5-hour window. Regular text chat stays uncapped on every plan, and the retired Thinking cap described below has not returned.

The short answer

What GPT-5 Thinking was, and why its cap surprised people

GPT-5.x on ChatGPT Plus ships in two modes: a fast default mode for everyday messages, and a slower reasoning mode called Thinking that works through harder problems step by step before answering. You can pick Thinking manually from the model picker, or leave ChatGPT on Auto and let it decide when a question warrants the extra reasoning time.

The surprise is structural. Plus users of the time already knew about the 3-hour rolling message cap on the default model - it was well documented and easy to find. What is much less visible is that Thinking runs on its own separate weekly allowance, with its own separate rolling window, and its own separate failure mode. A user who has been leaning on Thinking for a few days of hard debugging or analysis can watch it disappear from the model picker entirely, with no prior warning and no visible counter ticking down. Unlike the 3-hour cap, which most people learned about in their first week on Plus, the Thinking cap tended to surface only once someone genuinely relies on it - which is precisely when losing it hurts the most.

The numbers, by plan

OpenAI does not publish an exact figure for the Free plan's Thinking allowance anywhere in its help center - the 5-to-10-message range comes from independent limit-tracking guides that monitor the live rate-limit behavior, not from an official source. Treat it as a rough ceiling for planning, the same way you would treat any unpublished OpenAI limit. The Plus figure of roughly 3,000 messages a week was the one worth remembering: it was generous enough that most users never thought about it, until a multi-day stretch of hard analytical work runs the counter down faster than expected.

The rolling 7-day window, explained

Like the 3-hour default cap, the Thinking weekly cap was a rolling window, not a fixed calendar week. The 7-day clock started from your first Thinking message of the cycle - so two Plus users who both send 3,000 Thinking messages a week can have completely different reset days, based purely on when they each started.

The trap was the same shape as the 3-hour cap, just stretched over a week: burn through the budget early in the cycle and you wait longer for the reset, because the clock started counting the moment you first touched Thinking, not the moment you ran out. A user who sends their first Thinking message on a Monday and exhausts the 3,000-message budget by Thursday does not get a fresh allowance on Friday - they wait until the following Monday, when the 7-day window that started this cycle finally rolls over.

What counts against the weekly cap, and what quietly doesn't

This is the detail that generates the most confusion, and the reason this article exists: only Thinking messages you select yourself count against the 3,000-per-week budget. When ChatGPT's Auto mode decided on its own that a question deserved extra reasoning and silently routed it to Thinking, that message did not draw from your manual weekly counter at all. Auto-switching itself has since gone too: on 2026-09-14 OpenAI stopped Plus and Pro chats from switching automatically from Instant to a thinking level.

• Counts: any message sent while you have manually selected Thinking from the model picker, across chatgpt.com, the mobile apps, and the macOS desktop app.
• Did not count: messages where Auto mode silently upgraded a default-model conversation to Thinking-level reasoning behind the scenes.
• Did not count: the 3-hour default message cap of the time, covered separately in ChatGPT's rolling message window, explained - the two pools never overlap.

This split explains a pattern that shows up constantly in user reports: someone insists they have "barely used Thinking" and yet it has vanished from their picker. In most cases they have been manually selecting Thinking more than they realize, while also benefiting from Auto-routed Thinking answers they never noticed - and only the manual selections were ever going to count.

"Automatic switching from Instant to Thinking does not count toward the weekly manual limit." Paraphrased from independent ChatGPT Plus limit-tracking documentation, 2026

What actually happens when you hit the Thinking cap

Hitting the Thinking weekly cap behaved very differently from hitting the 3-hour default cap, and that difference mattered in practice.

In other words, running out of Thinking is a soft limit compared to the hard wall of the 3-hour cap. You lost access to the slower, more careful reasoning mode, but were never fully locked out of ChatGPT. The practical cost is qualitative rather than total: harder problems get answered by the faster default model instead of Thinking, with less step-by-step reasoning behind the response, until the weekly window resets.

How to check when your Thinking cap resets

ChatGPT does not give Thinking a running usage counter the way Claude's Settings > Usage page tracks its limits. The one built-in signal is available only after you have already hit the cap: hovering over or tapping the greyed-out Thinking option in the model picker surfaces the date it resets. Before that point, there is no visible indicator that the budget is running low - no warning at 90% used, no countdown, nothing in the interface hints that Thinking is about to disappear.

This mirrors the same transparency gap covered in our guide to how ChatGPT's limits really work in 2026: OpenAI communicates the cap at the moment you hit it, not before. For anyone doing sustained analytical or coding work across a full week, that means the only way to see the Thinking budget draining in real time is an external tracker that reads the session state continuously instead of waiting for the wall.

Frequently asked questions