How We Mapped the Usage Buckets

Dennis did not ask for another model comparison. He asked why Cursor showed 4% used after a day of factory work, while GrokBot had just burned another $300, and whether Cursor Ultra’s $400 was a loan on Grok. We drew the cash registers and published the map.

Name the confusion, do not paper over it

The live question was fifth-grade honest: Claude Cowork felt expensive. Codex felt cheaper. Cursor Grok 4.6 felt cheaper still. GrokBot the last few days did great work but had no effort picker and then a $300 extra-credit day. If that kept up it looked like $10,000 a month. Maybe Cursor was cheaper only because the limit is monthly?

Yes, Cursor’s included pools reset monthly. That is one real reason. It is not the only one, and it is not a reason to annualize GrokBot extra credits as if they were Cursor.

Three numbers were glued together that do not glue:

1. SuperGrok Heavy $300/mo is grok.com’s weekly buffet (Chat + Build + Imagine + Voice sharing one pool). It is also one door into Grok Bot. Linking Heavy can unlock Cursor Ultra at $0. The usage bars still do not mix.
2. Cursor Ultra $200/mo is a different company’s workshop. It rents Grok 4.6 in a Cursor Models pool, and it includes Grok Bot. You do not need grok.com if you live in Cursor.
3. Ultra’s $400 is the Other Models pool — Claude, GPT, Gemini inside Cursor at API prices. It is not extra Grok.

The fourth glue was the launch sticker: 2× included Grok 4.6 usage in Cursor and Grok Build for the first week after 12 Aug 2026, plus a 50% on-demand discount. A 4% bar during that week is not the forever bar.

The fifth glue, which we got wrong on the first publish: we treated Grok Bot as grok.com Heavy, and we said Heavy never fills Cursor. Grok Bot is the named-roster product announced 11 Aug 2026. Heavy’s four-at-once reply is an old one-chat trick. Desk SOP: /grok-bot/.

What we checked, not what we guessed

We used Dennis’s own 4% and $300 figures as he said them on 17 Aug. On 18 Aug we copied the live spending tab: Cursor Models 11%, Other Models 0%, Grok Bot weekly 100%, on-demand $358.49 / $380, 1.1B tokens for 12–18 Aug. Those receipts now sit on /skills-and-routines/. The SOP still tells every reader to open their own spending tab.

Tie it to the ladder we already teach

We already tell people a skill is a function and an agent is a person. We already tell them a job needs a schedule, a QA cycle, and working files. We already tell them batch nights belong on a local model. None of that said which cafeteria pays.

That hole is how you get a $300 GrokBot day after a week of good work: the workshop is right, the ticket is the weekly one with no effort knob, and extra credits look like “keep going” instead of grocery prices.

The map now lives at /usage-buckets/. The product split (workshop vs named roster vs four-at-once reply) lives at /grok-bot/. They sit next to persistent agents and the overnight worker. The older ChatGPT-to-Claude piece still stands as the 2026 quality test. This page is the 17 Aug money map, corrected 18 Aug when Dennis split Grok Bot from Heavy’s four agents.

Built 17 Aug 2026, corrected 18 Aug 2026, per the meta article prompt against /usage-buckets/ and /grok-bot/. Related: skill is how, routine is when · let the model pick the model · overnight worker · The System.