Over the past week Grok Bot subscribers kept finding the same problem in the same place: the weekly allowance was gone before the week was.
One user posted a meter sitting at 100 percent with four days left before the reset. Others reported the bar jumping from 50 to 100 percent overnight, or from 35 to 78, without any obvious activity. For a product sold on being always available, that is hard to read as a small change.
On November 1, Elon Musk answered a user who asked for a reset. His explanation was blunt: the cost of serving Grok Bot was unsustainably high relative to revenue. In a follow-up he wrote that reducing allowances was just being sensible, and that telling users about it was sensible too.
xAI's account splits in two halves. Some of the shortfall was a billing dashboard glitch; some was a deliberate adjustment to how usage is allocated. For anyone watching the meter, the distinction does not matter much, because the outcome is identical: the same money buys fewer tasks. Users on Reddit settled on a phrase for it, an indirect rug pull.
Where the cost comes from
It comes from the shape of the product. Every Bot runs on its own cloud computer, signs in to accounts, opens pages, moves between apps and works through multi-step tasks on a schedule. A chatbot does arithmetic once, when you ask. An agent works continuously in the background, and a single task can trigger a dozen model calls and web actions. The more useful it is, the closer the compute bill gets to linear, while the monthly fee stays flat.
Demand bent the curve sharply. In its quarterly results SpaceX disclosed that Grok Bot passed 800,000 paid weekly users in early October, roughly double the prior month, while daily users climbed from under 100,000 in early September to more than 600,000 by October 26. On the earnings call Musk called it the fastest-growing product he had ever seen.
The company is not short of money; the product's arithmetic is the problem. SpaceX reported $14.6 billion in third-quarter revenue, $7.56 billion of it from the Super Intelligence segment, or 52 percent. That segment produced $431 million in profit, weighed down by $1.2 billion in one-time stock-based compensation tied to the $60 billion acquisition of Cursor's parent Anysphere, which closed on August 14. Capital expenditure for the quarter was $19.6 billion, 84 percent of it aimed at AI infrastructure.
Put differently: the question is not whether the company can pay, but whether flat-rate pricing survives heavy agent use. The same GPUs rent for more to other labs.
This is not an isolated case
OpenAI tightened limits on flat-rate plans this year, which some users described as the end of an era. Anthropic blocked Claude Pro and Max subscriptions from powering third-party agent frameworks in April, citing complex agent sessions that can run $1,000 or more in API compute over a full working day, absorbed quietly for months. Meta's Muse, past ten million users, is very compute-intensive and not yet profitable, Mark Zuckerberg told investors.
Subscription pricing plus autonomous agents is turning into public arithmetic: the more customers use it, the more the provider loses. Usage caps are the only lever that closes the gap without raising the sticker price.
Three things to watch
First, whether xAI pushes heavy users toward per-task billing rather than simply tightening the cap. Second, whether there is a timetable for restoring allowances, which there is not. Third, whether the company publishes its measurement at all: how many tokens or tasks a weekly allowance contains has never been disclosed, which leaves buyers unable to judge what they are paying for.
For teams planning to put Grok Bot into a production workflow, the practical lesson is narrow. Run one well-defined, low-risk task for a full week, watch the meter, and only then pick a tier.
Grok Bot Weekly Caps Ran Out Early. Musk Says Serving It Costs More Than It Earns
← Previous: OpenAI's ChatGPT Wallet Lets Agents Pay, With a Human Still in the Loop
Next: →Grok Bot 1.0 Turns a Personal Agent Into a Shared Teammate in Slack, Teams and Calls →