AI chat usage limits compared: ChatGPT, Claude, Gemini and Grok
What ChatGPT, Claude, Gemini and Grok actually cap in September 2026, what each multiplier is measured against, and which numbers are still published.
Several major AI assistants advertise usage allowances as multipliers. Google sells 2x, 4x and 20x. Anthropic sells Max 5x and Max 20x. The multiplier goes on the plan page, and the unit it multiplies almost never does.
Which leaves a question that has become oddly hard to answer: if you pay, how much can you actually do before the app stops you?
Below is what each of the four biggest consumer assistants publishes, checked on 1 September 2026, with a date on every figure. Where a provider has stopped publishing numbers, this page says so.
The short version
Provider | Paid consumer tiers | What actually gets metered | Published figures? | Most recent change |
|---|---|---|---|---|
ChatGPT | Go, Plus, Pro | Files, images, voice, image generation. Text is uncapped | No, and text no longer needs them | Text message limits removed on all tiers, 6 August 2026 |
Claude | Pro, Max 5x, Max 20x | A five-hour session limit and a weekly limit, shared across claude.ai, Claude Code and Claude Desktop | No. The last official figures are from July 2025 and were never reissued | Weekly Claude Code limits change, 14 September 2026 |
Gemini | AI Plus, AI Pro, AI Ultra | Compute, which varies by prompt complexity, features and chat length | Multipliers only, and Google's own two plan pages disagree on Ultra's baseline | Gemini Notebook joins the same system, 2 September 2026 |
Grok | SuperGrok Plus, SuperGrok Heavy | Not disclosed in any source we could verify | No | Grok 4.6 released 12 August 2026 |
ChatGPT (OpenAI)
What gets capped. Not text messages. OpenAI removed text message limits on every ChatGPT tier on 6 August 2026, rolling out from the week of 10 August, and Free and Go got unlimited text chats. Separate limits still apply to file uploads, image uploads, voice and image generation.
The plans. Free, Go, Plus and Pro, alongside three business tiers. openai.com would not serve us a price for any of them. Ads appear for logged-in Free and Go users only. Every other tier is ad-free.
What is actually published. For text, nothing needs publishing. We could not confirm current per-tier figures for the four things still metered, so none are quoted. Any ChatGPT message cap dated 2025 or early 2026 stopped applying on 6 August.
What changed most recently. The same update made GPT-5.6 Luna the default model for Free and Go, both of which gained a Think button, and moved Plus and Pro to an updated GPT-5.6 Sol with a thinking slider.
Verdict. On text volume ChatGPT is the hardest of the four to run out of, because there is nothing left to run out of. A paid plan now buys model access, tools, uploads and an ad-free interface.
Claude (Anthropic)
What gets capped. Two meters run at once on every paid plan: a rolling five-hour session limit, and a weekly limit across all models resetting at a fixed time on your account. Anthropic's help centre says "your usage of all different Claude product surfaces (claude.ai, Claude Code, Claude Desktop) counts towards the same usage limit". A coding afternoon spends a writing evening's budget.
The plans. Free, Pro at $17 a month billed annually ($200 up front) or $20 monthly, then Max 5x at $100 and Max 20x at $200, both per the help centre.
What is actually published. Nothing numeric. No token, message or hour figure appears on the Max pricing page or its help article, whose footnote says only "Usage limits apply." The only concrete weekly figures Anthropic has ever published came with its July 2025 announcement, reported by TechCrunch: 40 to 80 weekly Sonnet hours on Pro, 140 to 280 on Max 5x, 240 to 480 on Max 20x. Thirteen months old, never republished, and still the source of almost every "Claude limits" number in circulation. Run them against the labels and Max 5x is about 3.5x Pro on the week, Max 20x about 6x, though that arithmetic is ours and not Anthropic's. The help centre qualifies it: 20x means per session, with the week counted separately. A class action from June 2026 alleges a shortfall in the same range.
What changed most recently. On 29 August 2026 Anthropic announced a permanent 25% rise in standard weekly Claude Code limits from 14 September, then said in the next post of the same thread that this is "a 17% reduction" against the boosted allowance. Both hold: a temporary 50% boost running since 13 May 2026 ends on the 13th. What Anthropic has not said, and its help centre does not answer, is whether the reduction comes off the shared pool or only a Claude Code share of it. If you chat rather than code, that is the part that matters. Anthropic also shipped Claude Fable 5.1 on 1 September 2026. Its help centre puts Fable outside plan usage limits on Pro and standard Team seats, running on purchased credits, while Max and premium seats may spend up to half their weekly limit on it. A plan can now name a model your allowance will not cover.
Verdict. Every meter here is documented except its size. The announced 14 September capacity is lower than the temporary allowance and higher than the pre-May base, and neither figure tells you how many messages you get.
Gemini (Google)
What gets capped. Compute rather than messages, since 20 May 2026. What a prompt costs depends on its complexity, the features it uses and how long the chat has run, so two prompts on the same plan can spend very different amounts. A five-hour reset window runs alongside weekly caps.
The plans. Free, Google AI Plus, Google AI Pro and Google AI Ultra, with Ultra sold at two levels. Google's own student offer page from August 2026 puts AI Pro's list price at $19.99 a month. No dollar figure is quoted for the other tiers: Google's plans page rendered no prices.
What is actually published. Multipliers and rules, never an absolute figure, and the multipliers do not agree with each other. Checked on 1 September 2026, Google's one.google.com plans page measures every tier against "people without a Google AI subscription": Plus 2x, Pro 4x, Ultra up to 20x. Its gemini.google subscriptions page measures Plus and Pro the same way but describes Ultra as up to 20 times higher "compared with the Pro plan", footnoting the two Ultra levels as 5x and 20x. That matches 9to5Google, which reported them as 5x and 20x Pro on 21 May 2026. Twenty times free and twenty times Pro are very different purchases, and Google has not reconciled them. AI Pro's 4x is also measured against the free plan, so a subscriber who had AI Pro before 20 May learns nothing about their own account.
What changed most recently. On 28 May 2026 Google capped how much quota one prompt can consume, stopped charging for failed requests ("If a request fails, you won't be charged. Our system mistakes are on us, not you.") and made Gemini 3.1 Flash Lite prompts free. From 2 September 2026 Gemini Notebook moves onto the same limits, resetting every five hours instead of daily, with in-product warnings and a "Generate later" option for heavy jobs.
Verdict. Google explains the mechanism better than anyone here and quantifies it worse. When a company's own two plan pages measure the top tier against different baselines, a subscriber cannot work out what they bought even in multiples. The May fixes were real, and pay-as-you-go credits point at billing for what you ask.
Grok (xAI)
What gets capped. Not disclosed. xAI publishes no consumer usage figures we could verify, so nothing here is quoted as a Grok limit. Same pattern as the rest: named tiers, no stated unit.
The plans. SuperGrok Plus and SuperGrok Heavy. Grok Bot, an always-on agent, left beta this year and runs on both. No price is quoted: x.ai and grok.com were both unreachable to us.
What is actually published. No numeric caps we could confirm. What is documented is a withdrawal: on 9 January 2026, amid the non-consensual imagery stories around the tool, X put Grok image generation behind a paid subscription. The clearest dated example of a Grok limit changing, and it was a safety decision rather than a capacity one.
What changed most recently. Grok 4.6 arrived on 12 August 2026 with a reported 500,000 token context window and four reasoning levels, per secondary coverage xAI has not confirmed anywhere we could reach. On 20 August 2026 Grok began returning word salad to direct queries on grok.com, which the official Grok account called "a rare temporary generation glitch". Availability is its own kind of limit.
Verdict. Grok is the least documented of the four on usage. Its plan pages let you choose on model access and features only, because the usage rules are not public. Treat a paid month as a test of your own habits.
What to check before you subscribe to anything
All four have converged on the same shape, so these six questions work on any plan page.
What is the multiplier measured against? Google says the free tier on one page and the Pro plan on another. Anthropic measures against Pro and, per its help centre, means the five-hour session window rather than the week. Same word, several different promises.
What is the unit? Hours, messages, tokens and compute all behave differently. A plan naming none of them has told you nothing. Only Anthropic ever published hours, in July 2025.
Is the allowance shared? Claude chat, Claude Code and Claude Desktop draw from one pool, and no plan page says so.
Is the number you like a promotion? Anthropic's temporary 50% weekly boost began on 13 May 2026 and ends on 13 September 2026. A boost you get used to reads as a cut when it goes.
Does the plan meter what you do most? Text is uncapped on ChatGPT now, files and images are not. If half your use is uploading PDFs, the unlimited headline is not about you.
Where do you live, and how old is the page? Google's support page for Gemini Spark, the agent it sells inside AI Pro and Ultra, says the agent works wherever Gemini apps do, "except in the European Economic Area, Nigeria, Switzerland, and the United Kingdom". A tier can include something your country cannot open. A page carrying no dates cannot be checked.
A disclosure, since we sell a subscription too. MultiChats puts 60 models from 18 providers behind one login, plans in the pricing section. Free accounts get 50 messages per 24 hours and two image generations a month. Our limits page publishes only the second, so the checklist catches us too.
Everything here was checked on 1 and 2 September 2026. Limits move quietly, so treat the dates as part of the facts.