Under a video called "Grok Bot usage just got nuked...", a commenter with 16 likes describes the wait that started this post: "I can't implement them today, 'cause I ran out of my usage two days ago, and have to wait 5 more days for my weekly usage to reset." (@jebediahkerman3946).
The replies under that video report the same problem: a plan that used to last a day now "evaporates in 7 hours" (5 likes); one commenter is at "70% in a day and a half." The complaints span vendors and plans, and they raise the same two questions: what is "usage," and why does it run out so fast?
A subscription doesn't buy unlimited AI. It buys an allowance that refills on a schedule, and some jobs use it up far faster than others. The distinction we'd carry through the rest of this post: a plan's limit is rarely one clock. It's a short clock and a long one, running at the same time.
What a usage limit is
A usage limit is the amount of work a plan lets you do before you wait for a reset, upgrade, or pay-as-you-go charge. It differs from the API, where developers pay per million tokens instead of drawing from an allowance; rate limits still apply on the API, but the bill tracks usage rather than refilling on a schedule. ChatGPT's help docs measure its consumer allowance in tokens (What Is a Token in AI? The Unit Behind Every Price and Limit explains the unit).
Reach the limit and the app pauses the work and shows when it resets. Both vendors describe the same options on the way through: wait for the reset, upgrade your plan, or pay as you go with credits (OpenAI's limits page lists all three; Anthropic's describes the first two in its limits guide and the third in its usage-credits page).
The two clocks
Claude's paid plans run a session limit that resets every five hours and weekly limits that reset once a week, per Anthropic's help center. The five-hour window starts when you send your first message, not on the hour. The weekly reset lands at a fixed day and time assigned to your account, the same every week regardless of when you subscribed. Anthropic's page also says usage across claude.ai, Claude Code, and Claude Desktop counts toward the same limit, and the Settings > Usage screen shows how much of the current session you've used, the time remaining, and your next weekly reset.
ChatGPT's current help page describes its allowance in tokens too: how much a task uses depends on the model you choose and the task you ask, and "more advanced models and complex tasks generally use more of your allowance." Its Settings → Usage screen shows combined remaining usage and reset times, and the page notes instant chat "will not draw from your usage." One change is coming: OpenAI says that starting in December, Plus, Pro, and Business plans will share one allowance across ChatGPT and Codex as the agentic Work capabilities come to the main app.
Vendors run their own clocks, and one vendor's empty allowance says nothing about the others. (How Many AI Subscriptions Do You Need? covers when a second plan beats an upgrade.)
What uses the most allowance
Three of these show up in the vendors' documentation; two come straight from the comment threads.
Agent work. A multi-step agent can make several model requests while it reads, decides, acts, and checks, and the docs agree that complex tasks use more of the allowance. Commenters report the effect: "direct messaging bots and doing research tasks eats up my usage fast."
Long conversations. A long chat costs more per reply than a fresh one, because the model processes the thread so far each time. Claude's help page says longer conversations that trigger automatic context management consume more of your usage limit, and recommends starting a new conversation when you approach the limit. A commenter under the personal-agent video described the same effect in a consumer bot: "every-time you talk to it reads the context of everything you ever said to it. Usage gets blown doing nothing."
Bigger models and higher effort. Claude's page names "the effort level you've selected" as a factor; OpenAI's names the model you choose. Routine jobs on the biggest model can spend allowance on steps that don't need it.
Scheduled routines. Commenters report routines as a silent cost: "Cron jobs and routines drain usage so fast!!" A routine can draw from the allowance whether or not anyone is watching its output, and old ones can keep running after you stop needing them.
Bots checking in with each other. The costliest report in the threads: one commenter's fleet of agents had spent "81% of my usage on communication with each other," coordination instead of work.
Five fixes that make an allowance last
Check the usage screen before you plan the work. Both apps show what's left and when it refills (ChatGPT: Settings → Usage; Claude: Settings > Usage), so a multi-day wait becomes a scheduled one.
Start a new conversation when one gets long. Claude's own documentation recommends it. Paste the short version of the conclusions into the fresh chat and stop paying to process the whole history.
Match the model to the job. Save the biggest model and the high effort setting for the work that needs them.
Audit the bots and routines. List what you've connected, find the ones you no longer use, and pause or delete them. (How to Keep Track of What Your AI Agents Are Doing walks through the audit.)
Plan heavy work around the reset. The weekly clock refills at a known time. Save the weekend project for right after it, and spread daily jobs so they don't all land inside one five-hour window. One commenter in the Grok Bot thread had collected "a bunch of optimisations" and was "waiting until Friday when my usage" refills; planning around the reset can make the same allowance cover more work.
The part vendors control
Providers can change an allowance after you've subscribed. Claude's help page states it directly: "We may adjust usage limits at our discretion to manage capacity and ensure fair access to all users." The Grok Bot thread shows what that looks like on the receiving end: commenters report their plans covering less work than before, and one writes that the provider offers "NO info on how it all works." The allowance is part of what a subscription sells, and it can change. The usage screen is the number to check today, and it's worth a look before you build a weekly routine on top of any plan.
FAQ
When does my AI usage reset?
The usage screen shows the reset times (ChatGPT: Settings → Usage; Claude: Settings > Usage). The two clocks work differently: the session window counts five hours from your first message, while the weekly reset lands on a fixed day and time assigned to your account, so it may not match the day you subscribed.
What's the difference between a usage limit and a context limit?
A usage limit controls how much you can use the service across all your conversations over time; a context limit controls how much one conversation can hold at once. Claude's help center draws exactly that line: quantity over time versus the depth and complexity of individual conversations. What Is Context in AI? Why It Makes or Breaks Your Results covers the context side.
Can I keep working after I hit the limit?
Usually, for a price. OpenAI's page lists buying credits beside waiting and upgrading. Claude's usage credits (which you enable first) switch your requests to pay-as-you-go pricing at standard API rates when an included limit runs out, and they stop at a monthly spend limit you set.
Do free AI plans have usage limits?
ChatGPT's free tier is the documented example: its FAQ describes everyday text chats as unlimited within abuse-prevention safeguards, while tools like image generation carry separate limits, and the app shows the reset time when you hit one.
Do all my apps share one usage pool?
Within a vendor, increasingly yes: Anthropic's page counts claude.ai, Claude Code, and Claude Desktop toward one limit, and OpenAI says ChatGPT and Codex will share one allowance on Plus, Pro, and Business plans starting in December. Across vendors, no: each company runs its own clocks.
Want to go further?
What Is Local AI? Running Models on Your Own Computer covers the option that skips the allowance entirely: routine jobs that run on your own hardware, with the cost moved to the machine instead of a subscription. The one-step version of this post: open the usage screen in the app you use most and read the two numbers, what's left and when it refills.