Over the Halloween weekend, people who pay for SpaceX's Grok Bot opened it and found a week's allowance already gone. Some reported going from half used to fully used overnight, without touching it. Elon Musk then wrote on X that the cost of serving Grok Bot had been unsustainably high compared with the revenue it brings in. He did not say what it costs or how far the allowances were cut.
An agent breaks the logic of a flat price
A chat is one question and one answer. An agent opens a page, reads it, moves to the next, drafts an email, checks it and logs in somewhere else. Every step runs through the model again with everything it has gathered so far, so an agent costs four to 15 times as much as a simple chat. It also keeps going when you are not watching.
A flat price works when most customers use a small part of what they pay for. With agents, the customer who gets the most value is the one who costs the most. A plan priced to cover 100 chat sessions a month would cover only about 7 to 25 agent tasks at those ratios. The subscription is from $30 a month, and the allowance is shared with Grok and with the coding tool Cursor, so a developer running long coding sessions draws on the same pool as someone sorting an inbox.
Growth makes this worse. When each new user costs more than they pay, a product that doubles its paying users in a month, as Grok Bot did in September, doubles the problem. A cut two weeks after Musk called it the fastest growth he had ever seen in a product is what you would expect.
SpaceX is not alone
OpenAI is moving the same way. From December, its Plus, Pro and Business plans share one token budget across ChatGPT and Codex, and harder tasks use it up faster. OpenAI is also cracking down on people who move between accounts to get around limits. When the Grok Bot first launched as an office agent, the pitch was a coworker that never stops. The bill for a coworker that never stops is what is now showing up.
What the phone plans suggest comes next
After AT&T and other carriers capped data, customers worked out which things used a lot of it, moved the heavy ones to Wi-Fi, and stopped worrying about the rest. Years later, unlimited plans returned, but with a slowdown for the heaviest users written into the fine print. Agents will probably follow the same path. For a while, plans will be allowances of work, with limits that move. Then, as each task gets cheaper to run, something that looks like unlimited will come back, with a ceiling you only meet if you are a heavy user.
That has a practical meaning for anyone who has handed a routine to an agent. Take a shop owner who has one chase late invoices and sort supplier email every week. Run the routine once and see what share of the weekly allowance it used. If it takes 5%, it is safe. If it takes 40%, one plan cannot carry it, and a quiet cut could stop it halfway through the week. Keep the instructions in a document you own, not only inside one app, so the routine can move to another agent when the allowance changes.