Blog Article
How Much Usage Does Grok Bot Use? I Ran 3 Large Tasks to Find Out
I cancelled SuperGrok Heavy, dropped to the X Ultra plan, and ran three large Grok bot tasks on camera. About 90 minutes of testing cost roughly 5-6% of my weekly
I cancelled my SuperGrok Heavy subscription.
Then I took the cheaper plan, gave three of my bot teams three rather large tasks, and watched the usage meter the whole time. If you've been reading the complaints on X about Grok bot usage being absolutely crazy, this is the test I wanted to see for myself.
Short answer: about 90 minutes of testing cost me roughly 5-6% of my weekly limit. The longer answer includes some output that was absolute trash and one build that impressed me.
Why I cancelled SuperGrok Heavy
It was my hope that SuperGrok Heavy, the legacy subscription, the top dog, would eventually carry some sort of benefit. After waiting a few weeks since they unified the X subscription, I don't think that's going to happen.
So I downgraded to the X Ultra plan. The reason is simple: Grok bot usage is unaffected between the two.
There is one caveat. On X Ultra, your Grok bot usage is shared with other programs in that plan. If you use Cursor, you would see a difference. That would have been the only benefit to me keeping SuperGrok Heavy.
I look at it as saving a hundred bucks a month, and I'm not complaining about that.
The starting line
Before I touched anything, I checked the meter:
- 12% used on my Ultra plan
- Resets in 4 days
One thing to know about that 12%: I don't do day-to-day work in Grok bot. That usage comes from the automations I have in place. At that pace, I wasn't going to hit 100% before the reset.
The three tasks
I like communicating out of my team chats, so each task went to a different team of bots. At least three bots get called per team.
1. The X team: short-form content from my live streams. I asked it to look at my past few live streams on the week and create short-form content from the most engaging topics. Hook first. I wanted at least 10 hooks, three core bodies, and three CTAs. I live streamed around 12 hours across Monday, Wednesday, and Friday, so that's a lot of data to go through.
2. The Clearmud team: 20 video ideas. Twenty variations of ideas for videos I can make about Grok bot, both educational and controversial. When those came back, I asked it to pick the top three and write out the scripts.
3. The engineering team: a video game. A realistic paintball first-person shooter. I don't use this team often. The last time I prompted them was about 2 months ago.
I kept my Google Workspace team out of it because there's client data in there.
I interrupted Grok and handed the build to Opus
When the engineering task kicked off, Grok started handling the build itself. I stepped in.
I told it Claude Opus 5.5 should handle the build, and then Grok and Codex should help fine-tune. I already know Opus 5.5 is the better of those models when it comes to building these types of things.
One wrinkle: it had to route the build locally because Claude Code on the remote computer was signed out. That's on me for not using engineering on the Grok bot VM in a while. Maybe it would have been a good idea to check that everything was authenticated before I hit record.
Usage check 1: only 3%
The ideation team finished first. The ideas it picked included effort level tests that don't cheat, and cheap models versus Claude or GPT on the same coding task, which was a community request. The scripts looked pretty solid, so I approved them.
At that point one of three tasks was done, with a lot of back and forth. I checked the meter [05:46].
Only 3%. That's not bad.
It's faster than I would like, sure. But outside of my ordinary automations, I'm still not going to hit 100%. And the engineering task was handed off to Claude, so that saved some usage too.
The short-form videos were absolute trash
Then the vertical videos came back from the X team.
We had a really good design system going for a while there, and all of these were absolute trash. I wanted this team to start clipping my long-form videos and live streams, because I record a lot of both. When I prompted the team to do that, it did not do that.
I'm not sugar coating things here. I'm not going to just show you the successes or the finely tuned outputs. I'm going to show you everything in between. No agent is perfect. No bot is perfect. And I need to spend more time on that specific team.
The second attempt was a little better, but way too zoomed in on me. What I want is a one-third split with the human on the bottom, two-thirds screen share on top, and text in the middle.
The teams that are dialed in
For contrast, my Clearmud team and my Google Workspace team are dialed in.
The Google Workspace team is making me money while I sleep. It qualifies leads for me. Whether it's inbound for work or inbound for sponsorships, the bots go back and forth in the conversation until it needs me, and then I come in and close the deal.
I wish I could show more of that, but there's too much sensitive data there.
Usage check 2: 5% total
At 1 hour and 15 minutes of recording, after several check-in nudges to the engineering team, I looked again [09:19].
A total of 5% used to record the video and run these demonstrations. One could argue that's not that much.
Here's my gripe, though. There's no transparency into the specific amount of Grok bot usage you get on each plan. Obviously the top plan gets more than the entry one. But between the middle and the top, I've been reading mixed community feedback. Some people think it's the same. I'd imagine the $200 a month plan gets the upper hand 10 times out of 10.
The paintball game works
Then the game was ready to play [09:49].
It did a pretty dang good job. The graphics, the bunkers, the shaders, the way you tap something with a paintball and it has this slight bounce. I was impressed.
It's technically not done yet. Claude was still running checks, and Grok fine-tunes after those finish. But it was playable.
My takeaway from that test:
- If you're not already handing off engineering work to the likes of Claude and Codex within Grok bot, you're doing it wrong.
- Grok doing all that heavy lifting does not compare to Opus 5.5 right now. Opus 5.5 is the more superior model.
- Grok 4.7 is still very impressive for agentic work, and for short-form content when it follows instructions.
The final numbers
By the end, the meter read 17%, up from the 12% I started with. In my wrap-up I put the past 90 minutes of testing at 6% usage on the week. Call it about 5-6% for the full session.
| Checkpoint | Usage | |---|---| | Before the test | 12% used, 4 days to reset | | One of three tasks done | 3% spent | | 1 hour 15 minutes in | 5% spent | | After the game and second video preview | 17% on the meter |
Based on all the complaints on X, I would have expected to burn 10% if not 15%. That's not the case, and we did quite a bit. Mind you, the engineering task did get handed off to Claude.
So don't believe everything you read on X, especially when somebody is trying to engagement rage bait you or engagement farm you.
Should you try Grok bot?
Grok bot is an awesome harness. If you don't care about self-sovereignty, if you don't care about hosting your own harness, go take a look at it.
This is not a paid advertisement. This is somebody who uses it daily. The two teams I rely on the most are the Google Workspace team and the Clearmud team. The X team I'm still fine-tuning.
Do you think 5-6% of a weekly limit is a lot for three large tasks? Tell me in the comments on the video, along with what you want to see me test next.
Watch the full test here: https://www.youtube.com/watch?v=W2NN6JlEJ4o
Watch the full walkthrough on YouTube.
Watch on YouTube