Most AI subscriptions no longer count messages. They meter a hidden budget that refills every five hours, caps out weekly, or both, and each vendor defines that budget differently. This post explains how the limits of Claude, Gemini, Grok and ChatGPT's Codex actually work according to their own documentation, what happens when you hit one, and which published numbers can and cannot be compared.
This comparison is based on official documentation, published specifications, and vendor pricing verified as of November 9, 2026. It is not based on hands-on use. No vendor publishes a message or token quota for these plans, so every comparison below is about structure and published multipliers, not raw allowance.
What each vendor says it measures
| Vendor | Unit | What drives usage, per the vendor |
|---|---|---|
| Claude | One pool shared by Claude.ai, Claude Code and Desktop | Conversation length and complexity, features used, model chosen and effort level; tools and connectors are token-intensive |
| Gemini app | Compute-based limits | Prompt complexity, models and features, chat length. Simple text chats cost very little, Deep Research and video generation cost the most |
| Antigravity | Agent work, not prompts | Simple tasks yield more prompts and complex tasks use more quota |
| Grok | One weekly pool across Grok products | Compute per action: a chat message uses little, video generation or long coding tasks use much more |
| Codex (ChatGPT plans) | Allowances per plan | Not described on the pages that could be read |
The details come from Claude's usage and length limits article, Google's Gemini Apps limits page, the Antigravity plans page and xAI's Grok FAQ, which says paid plans have used one weekly pool since June 2026 instead of separate daily limits per product. The practical meaning is that "number of prompts" is a poor predictor of when you run out. The same chat that costs almost nothing as plain text can cost far more if it triggers a long agent run or media generation.
Window anatomy: when the clock starts and what resets
| Item | Claude | Gemini app | Grok | Codex |
|---|---|---|---|---|
| Short window | Five hours, starting at your first message and resetting in full five hours later | Refreshes every five hours | None stated | Five-hour allowance on Plus and Business Standard |
| Weekly cap | Yes, on a reset day fixed to your account, not to when you started | Yes, you refresh every five hours until you reach it | One weekly pool, reset time shown in Settings, Usage | Yes; on the Pro tiers the weekly allowance is the covered one |
| Free plan | "Limited usage" | Standard limits, 32K-token context | Free-tier chat and voice limits, separate from the paid pool | Limited on Free |
Two structural points follow from those rules. First, a weekly cap sits on top of the short window, so using every five-hour window in a week is not possible. A week has about 33.6 five-hour windows (derived: 168 hours divided by 5), and the weekly cap is what binds you before that. Second, a Claude window opens when you send a message, so the timing of your first message of the day decides when the next full refill lands. Gemini's refresh and Grok's weekly pool have no equivalent start-on-first-use rule in the pages read.
The price-to-weekly-allowance ladder, normalized
Claude's pricing page now expresses every paid plan against Pro, which makes one clean normalization possible. Dividing each price by its weekly multiple gives these derived estimates, which measure vendor-defined usage and not tokens:
| Plan | Price | Weekly limit vs Pro | Derived price per Pro-weekly unit | Five-hour limit |
|---|---|---|---|---|
| Pro | $20 monthly ($17 annual) | 1x | $20 ($17 annual) | 1x baseline |
| Max | $100 | 5x | $20 | 5x Pro |
| Ultra | $200 | 10x | $20 | None for verified accounts |
The weekly price per unit is flat at $20, so moving up buys convenience and headroom, not a cheaper unit. What Ultra adds is the removal of the five-hour limit for verified accounts, which Claude's Ultra page says happens automatically based on account and payment history. The same page says Ultra is the renamed Max 20x plan, with existing Max 20x subscribers moved at the same price. Claude's earlier wording was "5x or 20x the usage of Pro per five-hour session," so the 10x weekly figure and the old 20x label are two different units. The pages do not say how they relate.
Google's ladder is anchored differently, as explained on its subscriptions page. Plus and Pro are multiples of the no-plan limit (2x and 4x), and Ultra is a multiple of Pro (5x or 20x). Chaining those gives about 20x and 80x of the no-plan limit, a derived estimate, and the context window also scales: 32K tokens with no plan, 128K on Plus, and 1M on Pro and Ultra. xAI's X Subscription states Premium at 4x, Super at 15x and Ultra at 50x of Plus, with the pool shared with Cursor.
What happens when you hit the limit
| Vendor | Options at the limit |
|---|---|
| Claude | Wait, upgrade (the new plan's limits apply immediately), or turn on usage credits billed at standard API rates, with a monthly spend cap you set and optional web-only auto-reload |
| Gemini app | Subscribers can keep chatting with Flash-Lite; otherwise upgrade or wait. Antigravity Pro and Ultra can buy AI credits, with a setting for Never or Always |
| Grok | Paid features pause until the weekly reset. You can buy Extra Usage Credits, priced higher per action than included usage, or turn on Auto Top Up |
| Codex | A free one-time banked reset, if OpenAI offers you one; otherwise wait |
Credits carry traps. Claude's usage-credit article says credits bought on or after December 7, 2026 expire 12 months after purchase, are non-refundable, and that Max's monthly API credits do not apply to them. Grok's credits expire one year after purchase unless stated otherwise. A Codex banked reset restores allowances below 100% back to 100% but does not refresh anything already full, and a reset can only be used when a covered allowance has under 80% remaining. OpenAI chooses who receives them, so holding an eligible plan does not guarantee an offer.
Where the sources conflict or go quiet
- Codex's five-hour limit: an OpenAI Support post dated July 28, 2026 says the limit was temporarily lifted for Plus, Pro and Business users as announced on July 12, with no end date. The help article on banked resets still lists a five-hour allowance for Plus and Business Standard.
- Pro tiers: the ChatGPT plans page says Pro is "5x or 20x," while the banked-reset article names Pro 100, Pro 200 and Pro 500. No price is given for Pro 500 on the pages read.
- Gemini credits: the subscriptions page says users can buy AI credits to extend limits, while the Gemini Apps limits page does not mention credits.
- Gemini Notebook: Google's notebook announcement says limits refresh every five hours "rather than daily" and does not mention a weekly limit, while the Gemini app page does.
- Claude Free: pricing says only "limited usage" and Claude's own help article gives no quota.
Who this system does not suit
- Anyone who needs a predictable number: none of the four vendors publishes a quota, and Anthropic says it may adjust limits at its discretion while Google says limits may change without notice.
- Free-plan users in busy periods: Google says users without a plan may find compute-heavy features like Deep Research unavailable during high demand, and their limits may be reduced before paid users'.
- Heavy users on a cap budget: Claude credits and Grok credits both bill above the plan price, and credits expire, so they suit occasional overflow rather than a standing habit.
- Teams planning around Codex: the temporary lift and the unreadable plan documentation make Codex limits the hardest of the four to plan around.
Which pattern fits which plan structure
| Your pattern | Structure that fits | Why, from the published rules |
|---|---|---|
| Short daily bursts | Five-hour window plans | A full refill arrives within hours, and the weekly cap rarely binds |
| One long, uninterrupted work session | Claude Ultra | No five-hour limit for verified accounts, weekly cap only |
| Irregular heavy weeks | Grok's weekly pool or banked resets | Unused allowance is not split into short windows, and Codex resets can restore a covered allowance |
| Occasional overflow past the plan | Usage credits | Billed at API rates, and you cap monthly spend yourself |
| Long documents in one chat | A plan with a large context window | Context limits are separate from usage limits: Gemini gives 32K, 128K or 1M by plan, and Claude's newest models reach 1M tokens on paid plans |
FAQ
Why do AI plans reset every five hours?
Vendors describe the budget as compute-based and refreshed in short windows, with Google saying five-hour refreshes let users keep working through the day. Weekly caps then limit total use across the week.
What is the difference between a usage limit and a length limit?
A usage limit caps how much you can use a service over a period. A length limit caps how long a single conversation can be, based on the context window, as Claude's help article explains.
Does Claude Ultra have no limits?
No. Verified Ultra accounts have no five-hour limit, but they still have a weekly limit across all models and a separate weekly limit for Fable.
Last verified: November 9, 2026. Limits, plan names and promotions change often, and the Codex five-hour limit was temporarily lifted at the time of checking. Pages that could not be read: OpenAI's Codex pricing page and its help article on using Codex with a ChatGPT plan.
Sources
- Claude Help Center — How do usage and length limits work?
- Claude Help Center — Manage usage credits for paid Claude plans
- Claude Help Center — What is the Ultra plan?
- Anthropic — Claude pricing
- Gemini Apps Help — Gemini Apps limits and feature availability
- Google — Gemini subscriptions
- Google Blog — Flexible usage limits for Gemini Notebook
- Google Antigravity — Plans
- xAI Docs — Grok website and apps FAQ
- xAI — Introducing The X Subscription
- OpenAI Help Center — How resets work in Codex
- OpenAI Developer Community — OpenAI Support reply on the Codex five-hour limit
- OpenAI — ChatGPT plans page