Claude vs ChatGPT Limits, Side by Side in 2026

In late 2026, Claude runs two rolling caps on Pro (a 5-hour session and a 7-day weekly window covering every model) and three on Max (those two plus a weekly meter scoped to Fable). ChatGPT dropped its text-chat message caps entirely on 2026-08-06 - what remains limited are images, files, voice, Deep Research, Agent mode, and (on Plus, since 2026-08-26) rolling 5-hour Codex and ChatGPT Work windows. Different shapes, different ways they bite. The structural difference that matters more than any cap's exact size is transparency: Anthropic publishes your remaining percentage, OpenAI does not. This article puts both providers side by side, with sources.

Update, 2026-09-19: Max plans no longer run a separate weekly limit scoped to Opus. Opus and Sonnet are both current, fully available models, and both draw from the single weekly limit that covers every model. The second weekly meter on Max is scoped to Fable instead: Max includes Fable for up to 50% of the weekly limit, after which Fable keeps running on pay-as-you-go usage credits. Pro has one weekly limit across all models and bills Fable from usage credits from the first message. Checked against a live Claude Max 5x account that used Fable, Opus and Sonnet in the same week: Settings > Usage shows current session, this week, and Fable this week, with no Opus meter.

Update, 2026-09-18: On the Claude side: Anthropic's Fable models (Fable 5, and Fable 5.1 since 2026-09-01) sit above Opus. On Max plans they are included for up to 50% of your weekly limit at no extra cost, and they drain that limit faster than Opus or Sonnet; past the 50% mark Fable keeps running on pay-as-you-go usage credits. On Pro, Fable is not part of the plan limits at all: every Fable message is billed from usage credits. The Free plan has no Fable access. On the ChatGPT side: 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.

Claude vs ChatGPT limits, in one sentence

Claude's limits are windows you can see (5-hour session, 7-day all-models weekly, and a Fable weekly meter on Max); ChatGPT's are a handful of windows you cannot see - and since 2026-08-06 none of them apply to regular text chat, which is uncapped on every plan. What OpenAI still limits: images, files, voice, Deep Research, Agent mode, and the 5-hour Codex/ChatGPT Work windows on Plus. For chat, ChatGPT is now effectively unlimited; for heavy agentic coding, Claude's visible caps against ChatGPT's invisible Codex windows is the real comparison. The biggest practical difference remains that Anthropic shows you your percentage and OpenAI does not.

The rest of this article is the long version: the comparison table, what each cap actually does, the transparency gap, five realistic workflow scenarios, the cost-per-hour math, and the decision matrix for picking which provider to push first.

The big side-by-side table

Every active rate-limit window across both providers' consumer plans, with the source for each. Hour counts are rough because both providers express budgets in usage percentage rather than absolute hours, and the conversion depends on what kind of work you do. The structure of each cap is stable; the absolute numbers shift when the provider adjusts limits (which Anthropic did mid-2025, OpenAI does quietly).

For the underlying mechanics on the Anthropic side see how Claude's usage limits really work in 2026. For the OpenAI side see how ChatGPT's limits really work in 2026. This article is the comparison; those two are the deep dives.

Claude's three caps, in plain English

Claude's consumer-tier limits are the most transparent in the industry as of 2026. Both Pro ($20) and Max ($100/$200) run the same three-window structure; Max just raises the budget at each window.

The 5-hour session window

Every active Claude conversation lives inside a rolling 5-hour window. It starts the moment you send your first message and resets exactly 5 hours later - not at midnight, not at any clock time. While inside that window, every model call (Sonnet, Haiku, and on Max plans Opus) deducts from your session budget at a weighted rate. The web Settings > Usage page and the /usage slash command in Claude Code show your remaining percentage in real time. Anthropic's help-center article on Pro usage confirms the rolling-window math.

The weekly all-models window

On top of the 5-hour session window, Claude added a 7-day weekly cap in mid-2025 to control sustained heavy use. It tracks all models combined and resets 7 days after your first message of the current week. Hitting the weekly cap on a Wednesday afternoon means waiting until the following Wednesday afternoon to recover - not Sunday-night-midnight as some users assume. This is the cap that bites Claude Code users hardest because agentic runs can chew through it without it feeling fast.

The weekly Fable meter (Max only)

Max plans (5x at $100, 20x at $200) add a second weekly meter, scoped to Anthropic's Fable models. Fable is included on Max for up to 50% of your weekly limit; past that ceiling Fable keeps working on pay-as-you-go usage credits while Sonnet and Opus carry on from the weekly pool. Pro has no Fable meter because Pro bills every Fable message from usage credits from the first one.

Retired: Max plans used to run a third cap scoped to Opus only, which silently downgraded you to Sonnet when it ran out. That meter no longer appears, and Opus usage counts against the all-models weekly cap like every other model. Opus itself is unchanged and fully available.

ChatGPT's caps, in plain English

On 2026-08-06 OpenAI removed message caps from text chats on every plan - regular conversations are now uncapped, on GPT-5.6 Sol for Plus and Pro and GPT-5.6 Luna for Free and Go. What remains limited: images, file uploads, voice, Deep Research, Agent mode, and since 2026-08-26 the rolling 5-hour Codex and ChatGPT Work windows on Plus (roughly 10-100 Sol, 25-200 Terra, or 250-2,000 Luna messages per 5 hours; Pro gets 5x-20x those ranges and is exempt from the restored Work limit). The two retired caps below are kept for reference.

The rolling 3-hour window (Plus) - retired August 2026

ChatGPT Plus ran a rolling 3-hour message-count cap. Approximately 80 messages per 3 hours to GPT-class models was the widely-reported number as of mid-2025 (later ~160), but OpenAI deliberately left it undefined and adjusted under load. The window was rolling: 3 hours after your first message of the current cycle, your budget refreshed.

The weekly Thinking cap - retired August 2026

Plus and Pro both ran a weekly cap on Thinking-mode messages, retired along with the manually-selected Thinking mode itself (reasoning is now a thinking-effort setting). The cap was meaningfully smaller than the 3-hour count on Plus, and most Plus users hit it within a long debugging session. Like the 3-hour window, OpenAI never published the exact number.

Pro's "fair use" ceiling

ChatGPT Pro at the higher price tier is marketed as effectively unlimited GPT-class usage, with a vague "fair use" qualifier. In practice there is a daily soft ceiling - measured in thousands of messages, hit only by automated workflows or near-continuous use. The exact number is internal. For non-agent workflows, Pro is functionally unlimited.

The transparency gap, and why it matters more than the cap shape

The most important thing about these two limit systems is not their relative size. It is that one provider tells you your remaining percentage and the other does not.

On Claude, you can open Settings > Usage on claude.ai or run /usage in Claude Code and see your exact percentage remaining on each of the three caps, plus the time-until-reset on each. This makes pacing possible: when you see your weekly drop to 18% on Tuesday afternoon, you know you need to be selective for the rest of the week.

On ChatGPT, no such surface exists. There is no Settings > Usage page. There is no /usage command. There is no API endpoint OpenAI publishes for your remaining quota on the consumer plans. The first signal that you've hit a cap is the cap itself - the message "you've hit the limit, please try again at 6:47 PM" or similar. You cannot pace what you cannot see.

This is a real product difference and it persists in 2026 with no public timeline for OpenAI to change. Anthropic's choice to publish percentages is itself a marketing position - they bet that informed users are happier users, even if the number is lower than expected. OpenAI's choice not to is also a position - that opacity reduces user anxiety in aggregate even though it bites the heavy users.

The third-party menu-bar tracker class of tool exists in large part to fill this gap on the OpenAI side. Tools like tokn.watch read OpenAI's internal rate-limit endpoint via your existing browser session and surface a remaining percentage the official UI never shows. The Claude side is already transparent, so the same tool reads Anthropic's official endpoint. Either way you end up with a glanceable percentage in your menu bar regardless of which provider you're using.

Five realistic scenarios - which provider bites first

The cap structures don't translate directly to "which is better". They behave differently for different workflows. Five common shapes:

The $/hour math, both sides

The same back-of-envelope as the Claude Pro vs Max piece, extended across both providers. Hours represent realistic intensity, not theoretical max throughput.

Two things to read off this table. First, at heavy intensity, Claude Pro $20 is the best $/hour deal at the entry tier because its caps are larger than ChatGPT Plus's at the same price. Second, ChatGPT Pro at $200 is cheaper per hour than Claude Max 5x at $100 - because Pro is functionally uncapped at non-agent intensity, while Max 5x leaves headroom for agentic runs that most non-agent users don't actually need.

So the upgrade story is not "Claude wins" or "ChatGPT wins" - it is "$20 Claude beats $20 ChatGPT for heavy code; $200 ChatGPT beats $100 Claude for everyday heavy use". Pick the price point and the workflow first; the provider falls out of that.

The decision matrix - which provider to push first

If you're not sure which subscription to commit to, use the following matrix. It is opinionated but holds for most readers.

• Code-heavy workflow, Claude Code or Cursor primary. Start on Claude Pro $20. Measure for two weeks. If you hit the weekly cap before Wednesday twice in a row, upgrade to Max 5x.
• Conversational / chat-heavy workflow, multimodal occasionally. Start on ChatGPT Plus $20. Text chat has no cap at all, so sparse chat use never runs into a limit.
• You want a meter on both providers regardless of which you commit to. A menu-bar tracker shows both percentages in real time, so you can see when one is getting close and shift to the other - effectively a manual load balancer. tokn.watch is the only Mac menu-bar app that surfaces both providers in the same item.
• You use both providers heavily. Run dual $20 subscriptions; this is cheaper than a single $100 upgrade and gives you the strengths of both models.
• You're building agents. Both providers' subscription tiers are the wrong product class for you. Move to the respective APIs (pay-per-token), where there are no rolling-window caps. Use the subscriptions for the human-driven parts of your workflow only.
• You hit limits every week and are not building agents. If on Claude → upgrade to Max 5x. If on ChatGPT → upgrade to Pro. Don't switch providers reactively; that just resets the learning curve. Upgrade within the provider you're already using.

For the Pro / Max plan-choice question on the Claude side specifically, the Claude Pro vs Max decision guide covers the math in depth. For the Plus / Pro question on the ChatGPT side, see how ChatGPT's limits really work in 2026.

What this article does not try to settle

To be clear about scope: this is a limits-mechanics comparison. It does not try to answer which model is smarter, which is better at code, which is funnier, which is better at images, which is more aligned, or which company has the better roadmap. Those are real questions, important questions, and outside the scope of this piece. The Claude-vs-ChatGPT capability comparison space is well-served by other publications (Zapier has a good one); the limits-specific gap is what tokn.watch's own positioning is built around.

The honest reason limits matter more than they should in 2026 is that both providers' subscription tiers leave heavy users at the edge of those limits most weeks. If subscriptions were over-provisioned the way mobile-data plans are now, none of this would matter. As long as Anthropic and OpenAI sell limit-priced subscriptions, the limits comparison is the comparison that determines your actual experience.

Frequently asked questions