Musk says Grok Bot cost too much to run as users hit weekly limits early

Elon Musk has confirmed that SpaceX deliberately reduced how much people can use Grok Bot, its AI agent, because the cost of running it had grown "unsustainably high relative to revenue". He gave the explanation in a reply on X on 1 November, after a weekend in which paying users found their weekly allowances suddenly used up.

The admission matters well beyond one product. Agents that operate a computer on a user's behalf burn far more compute than a chatbot answering questions, and nobody in the industry has yet shown a pricing model that clearly covers that bill.

What changed over the weekend

Over the Halloween weekend, a number of Grok Bot subscribers posted on X that their weekly usage meters had jumped, in some cases from around half used to fully used, without them running any new tasks. Many assumed it was a bug.

It was not. Musk said in a series of replies that the cost of serving the agent had outrun its income, and described cutting the allowance and telling users about it as simply sensible. To one user who complained, he asked whether moving to a higher-priced plan was really too expensive.

As of Monday, Musk had not said what the new limits are or whether they will change again, and he had not addressed the change in a post on his own main feed.

A product sold as a growth story

The timing is awkward. Grok Bot launched in August and is sold in tiers that run from free up to $200 a month, each with a weekly allowance that resets. Weekly caps are now common among AI agents; Meta uses them for Muse, its own personal agent, to keep compute bills under control.

SpaceX, which went public this year and now houses the Grok business, had been presenting the agent as a success. In its most recent earnings, the company said Grok Bot had 800,000 paying weekly users, and Musk told investors it showed "the fastest growth of a product that I've ever seen in my life". He also said he expected the agent to end up using most of the company's AI compute.

SpaceX said in a filing last month that AI now accounts for 52% of its income, and that its renamed "Super Intelligence" division made a profit of $431m. That figure would have been higher without $1.2bn in one-off costs tied to its purchase of Cursor, the AI coding tool.

The company has also been widening what Grok Bot does. Musk said in October that the agent would draw on models from other companies, including Anthropic's Claude Opus 5.5, Midjourney and Suno, and SpaceX plans to bundle it into a new $8-a-month X subscription alongside X Premium and Cursor.

Why agents are hard to price

An AI agent of this kind runs a virtual computer, opening pages, clicking and filling in forms to complete a task such as buying a present. Every step is a model call. A single errand can cost many times what a chat reply costs, and heavy users can consume far more than a flat subscription brings in.

That puts providers in a bind. Generous allowances win users but lose money on the most active ones; tight caps protect margins but frustrate the people most likely to pay more. Meta has not said what Muse costs to run, though Mark Zuckerberg told investors last month that the agent is very compute-intensive and not yet profitable.

Some users were quick to point out that Musk recently argued saving money would soon be pointless in an age of AI abundance. For now, at least, his own agent is being rationed.

What European builders can take from it

For European startups building agent products, the episode is a useful warning. Usage-based costs can rise faster than subscription revenue, and changing limits without notice damages trust with exactly the customers a company most wants to keep. Clear allowances, published in advance, and plans that scale with real usage are likely to age better than open-ended promises.

The open question is what Grok Bot's limits will look like once SpaceX settles them, and whether the cheaper bundled X subscription can carry an agent its own chief executive says is expensive to run.