I run bot.ski with a small team of Grok Bots. One builds the site, one checks it, one handles deploys, and a few more do research and writing. I'm on a SuperGrok Heavy subscription linked to Grok Bot, and that subscription is what gives me my Grok Bot usage.
Here's my problem. Every week when my usage resets, I seem to hit my weekly Grok Bot limit faster than the week before. These numbers are my own experience, not a measurement anyone else should expect. Your plan, your Bots, and your jobs are different from mine.
In mid-September I was running out about four or five days before the next reset. On September 13 I posted that I kept hitting my limit early, “Even with Super Grok Heavy and Cursor Pro +.” I learned later that those two don't add up: the official Plans and billing page says a Cursor plan and a SuperGrok link do not stack.
The next day I joked that, since hitting the Grok Bot limit, I'd also maxed out my GPT, Gemini, and Claude. It was funny, and it was also true.
On September 17 I was waiting on a reset that hadn't come yet, still partway through my week.
By September 23 I got a usage warning with most of the week still ahead of me, and I asked out loud, “Am I overusing it?”
Then last week I hit the limit in a little over a day, about 40 hours after the reset. That's when I stopped guessing and went looking. This post is what I found and what I changed.
What my usage screen shows
When I open my usage screen, I see separate SuperGrok Heavy meters for things like Chat, Automations, App Builder, Imagine, and Voice, and then a separate Weekly Grok Bot Limit. That one is the meter that runs out on me.
That lines up with the official Plans and billing page. It says Grok Bot usage is metered on your Cursor account, not on Grok or X, and that a SuperGrok link grants Grok Bot usage without creating a second meter on Grok or X. So the Grok Bot meter is its own thing, and the other SuperGrok meters don't tell you how much Grok Bot you have left.
What my audit found
On September 30, after an ordinary hour of work ate a surprising chunk of my week, I stopped and had my Bots audit what was using the most. There's no per-Bot meter on the computer, so this was a rough audit based on the size of what each tool returned. It still pointed straight at the problem.
The biggest users, roughly in order:
- Page snapshots in a loop. One of my Bots was waiting on a slow outside agent in a browser tab. Every time it checked, it captured the whole page again. Several of those captures were seconds apart and almost identical.
- Whole repository listings. Two of my Bots pulled a full listing of my code project, twice in a few minutes, and one pulled an entire large file to change one entry.
- Long chat histories and a stale to-do list. My main Bot's task list had grown past a hundred items, most of them long finished, and it got re-read on every reply. Each helper Bot carried its own long history too.
- Full-page QA screenshots, even when every check passed.
- A full docs scan. One Bot fetched the whole official docs site once, and a new daily check I'd just set up would have done that every day.
What wasn't the problem: my house rules. I was sure all the briefing text I send my Bots was the cost. It turned out to be tiny next to what the tools were pulling in.
What I cut and paused
That night I changed six things:
- Patient checking. When a Bot waits on something slow, it does one small check every few minutes and captures the full result once, when it's done.
- Diffs, not trees. My Bots fetch the file they need by path, or the diff between two commits, never a whole repository tree.
- Trimmed context. The to-do list went down to about fifteen live items. Briefs now point to one shared rules skill instead of pasting the rules every time.
- A local docs cache. I keep a local snapshot of the docs with each page's last-updated date from the sitemap. The daily docs check reads the sitemap and re-fetches only pages that changed.
- Screenshots only on failure. The visible screen only, logged as text when a check passes, and opened only when something fails.
- Paused failing routines. Four scheduled routines had been failing and retrying. I paused them, not deleted them, until they're fixed.
The handoff lesson, a day later
On October 1, I found another one. My builder Bot ran the tests on every change. Then my QA Bot, Keep, rebuilt the site and ran the full test suite plus browser checks at several screen widths on every pull request. After each deploy, my deploy Bot, Mr. Laser, reran the tests, checked every link, and loaded several widths again. The same checks were running three times. A full QA pass is one of the heavier runs my Bots do, so that added up.
Now Keep only reviews code and layout changes. Content-only edits get a quick check from my builder Bot and a look from me. After each deploy, Mr. Laser does a quick smoke test: the site is up, the changed page loads, and key settings are still off. The full sweep only runs for nav, layout, or config changes.
If your Bots hand work to each other, look for the same check at each handoff. Keep each check where it catches something nothing else can.
I'm not the only one
Plenty of other people have posted about running into their limit. These are their own reports, in their own words, not measurements I've checked.
AC Thompson (@acxsasx) hit his Grok Bot limit for the first time on September 30 and described it as “Stone cold silence.”
James Cornman (@jizaymes) upgraded to SuperGrok Heavy and said his usage started moving again within minutes, even with no Bots or routines running, and asked what he was doing wrong. For what it's worth, the plans page says it can take up to 24 hours for weekly usage to move to a new tier after you upgrade a linked SuperGrok plan.
Chad Moran (@ChadMoran) said one task he tried in Grok Bot used a big share of his weekly limit and failed, while Muse got it done using less. That's his comparison on one task, not a benchmark, but it's a fair reminder that a failed run still uses usage.
ChristCast (@christcast) posted on September 24 that, for the first time, they were about to hit their SuperGrok Heavy limit with a lot of planned updates, polish, and reviews in the queue.
And Bek (@miniyarov) said the Grok Bot weekly usage that comes with Cursor Pro runs out quickly for him. The official plans page lists Cursor Pro as weekly usage below Pro+, and Pro+ below Ultra, without giving amounts.
What I wish existed
On September 21 I floated an idea: once you've hit both the weekly limit and your on-demand limit, let your Bots keep going at a much slower speed until the next reset, instead of stopping. That's my wish, not a feature. Today, the official docs say Grok Bot stops when weekly usage runs out and on-demand is off.
Tips that are working for me
These work in my setup. Take what fits and skip what doesn't.
- Measure before you guess. Ask your Bots what they pulled in recently, biggest first. The biggest items usually point at the burn.
- Check on slow things occasionally. One small check (a status line, a title) every few minutes beats rapid-fire page snapshots.
- Brief once, then point. Put your house rules in a skill and tell helpers to follow it, instead of pasting long rules into every message.
- Fetch single files or a diff. Never pull a whole repository tree to read one line.
- Read docs from a local copy. Compare the sitemap's last-modified dates before you re-fetch anything.
- Prefer webhooks or deltas over heavy daily scrapes. A routine that fires on an event, or fetches only what changed, does less work than one that re-reads everything on a clock.
- Keep freshness routines quiet when nothing changed. A watch routine should say nothing and do as little as possible on a no-change day.
- Pause failing routines. A routine that fails every run still uses usage on every attempt. Pause it, fix it, then resume it.
- Keep context short. Close finished tasks and trim standing lists. Anything a Bot carries gets re-read on every turn.
- Wake other Bots sparingly. Send FYIs as non-urgent notes so nobody spends a turn just to say "got it."
- Look for repeated checks at handoffs. See the lesson above.
- Move heavy coding off the Grok Bot meter. I do token-heavy coding in Grok Build or Cursor on my own PC, so in my setup it doesn't come out of my weekly Grok Bot limit.
- Check your usage before big jobs. The official Plans and billing page suggests scoping tasks tightly and checking usage on the plan screen before large or long agent jobs. It says that about the free trial, and I find it works just as well for the weekly limit.
For more on writing tighter jobs, see tighter instruction examples. For Bots on one shared computer, the Botski virtual machine tips cover screenshots and files. And meter is the word we use around here for the weekly budget.
How Grok Bot limits work (from the official docs)
This part comes from the official docs, not from me. I'm not printing prices or usage amounts, because the docs don't publish them as numbers.
- Usage resets weekly. The Grok Bot overview says usage resets weekly, and the FAQ says Grok Bot subscriptions include weekly usage.
- Plans differ in size, not in kind. Plans and billing describes SuperGrok Heavy as the highest linked usage, and Cursor Ultra as the highest Cursor weekly usage, without giving amounts.
- Plans don't stack. The same page says a Cursor plan and a SuperGrok or X Premium+ link don't add usage on top of each other. The FAQ words it differently: if you have both, Grok Bot uses whichever has more usage. Our plans and add note covers that disagreement.
- On-demand is the way past the weekly limit. When weekly usage runs out, extra usage can continue as on-demand usage if it's enabled, billed through Cursor and counted toward your on-demand monthly limit (Plans and billing).
- If on-demand is off, Grok Bot stops. The screen says you've reached your Grok Bot usage limit, and it resets with weekly usage (Plans and billing).
- A monthly limit isn't a mid-run stop. A Bot already working can finish past the on-demand monthly limit (Plans and billing).
- Where to look. In Settings, Usage & Billing shows weekly included usage and on-demand usage for eligible accounts, and the account menu can show Weekly usage at a glance (Settings and notifications).
- No model picker. The same page says Cursor manages model selection, so switching to a cheaper model isn't a setting you can change.