Grok Bot's usage limits: Musk says the cost was 'unsustainably high'
On Sunday a subscriber asked Elon Musk on X for a reset of his Grok Bot weekly allowance, saying a lot of people had run out "really fast this week." Musk's reply was one sentence long: "The cost of serving Grok Bot was unsustainably high relative to revenue."
He gave no numbers. He did not say whether SpaceXSI had cut allowances or would restore them, and Forbes, which broke out the exchange, noted both absences. In a second reply he asked a user whether the jump to a higher-usage plan cost too much, and in a third he wrote "just being sensible," which Business Insider read as a defense of smaller allowances rather than an apology for them.
That is the first time the boss of a frontier lab has said in public what every metered agent product implies: the seat is priced like software and delivered like compute.
The $90 bundle that became a $30 one
Grok Bot is not the Grok chatbot. It is an agent with its own virtual computer that signs into your apps and works through multi-step jobs — shopping, bookings, forms — and it launched in beta on August 11.
On October 12, SpaceXSI folded it into The X Subscription. The company's own announcement priced the entry tier at $30 a month and described it as "the benefits of SuperGrok, Cursor Pro, and X Premium+, previously $90/month combined." Super is $100, Ultra $200, and every paid tier draws on "a generous pool of Grok model usage shared across Grok and Cursor."
Three weeks separate that announcement from Musk's admission. The bundle took a third off the price of three products and added an always-on agent to the deal; the agent is the part that spends money continuously.
There is also a number that makes the growth look better than the economics. On SpaceX's October 20 earnings call, president Gwynne Shotwell said Grok Bot had passed 800,000 paid weekly users in early October, more than double the month before, as The Next Web reported. Eight hundred thousand people is a product. It is also 800,000 weekly allowances to fund.
What the meters did over the weekend
The complaints were specific enough to check. The Telegraph found a user whose meter went from 33% to 66% with no change in activity. On Cursor's community forum, an Ultra subscriber watched the weekly meter flip between 77% and 100% in real time; another reported a jump from 34% to 73% overnight, with their exported usage CSV showing no recorded activity between 19:01 and 07:05. One user asked their own Bot what the number was and got a different answer from the app.
Cursor's status page logged a "billing and usage dashboard" degradation on Sunday, affecting cursor.com and the enterprise dashboard, later marked resolved. Forbes reported it is unclear whether that incident caused the readings. It matters for the diagnosis: some of the spikes were probably a dashboard bug, and some were probably the allowances themselves.
Musk did not distinguish between the two. Neither did SpaceXSI.
What a Bot actually costs, and why the price list never said
Grok 4.7's list rate is $2 per million input tokens and $6 per million output tokens below a 200,000-token prompt. Do the arithmetic on the entry tier: $30 a month buys 5 million output tokens at list price. An agent that browses screenshots pages, re-reads them, retries failed clicks and keeps its own context alive spends tokens on steps, not answers. A chat session that feels expensive is a rounding error next to an afternoon of computer use.
Then the cost base got worse on purpose. On October 7, Musk said Grok Bot would route work to "the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs," and a Grok Bot team member said Opus 5.5 was rolling out to power all Bots. That is a good product decision — you send the hard reasoning to a rival's model — and it means paying Anthropic's rates for the tasks your own models handle worst, inside a seat priced against Grok's own cost base. SpaceXSI publishes no separate charge for Bot usage of those models; Cursor's third-party models are billed apart, Grok Bot's are not described that way.
The billing structure compounds it. One pool covers Grok and Cursor, so a heavy Bot week is a coding week you don't get. This morning OpenAI moved in the same direction, merging ChatGPT's chat and Work tabs and putting ChatGPT and Codex on a single allowance pool, as TechCrunch reported — with the same warning that some users will hit resets more often.
This is not one company's problem. Mark Zuckerberg told investors Meta's Muse is "very compute-intensive" and not yet profitable at 10 million users. OpenAI stopped taking new $200 Pro signups in September, citing demand. The Telegraph's framing for the weekend was blunt: AI firms ration usage as losses mount.
Who wins, and who eats the difference
The winners are meters. On-demand billing, per-task pricing, and enterprise seats with pooled allowances and an admin who can see the burn all get stronger when the alternative is a public admission that the flat fee doesn't cover the cost. Enterprises were always going to want this; now consumer products are being pushed toward it by arithmetic.
The losers are the "unlimited agent" pitch and the users who bought it. But the deeper damage is to trust, and it is not about the cap. A meter that can move 40 points overnight while your own log shows no work destroys the one thing a subscription sells: the ability to forecast a week. Teams on Cursor's forum said they stopped all work at a false 100%. That is a worse outcome than a smaller allowance honestly labeled.
We have been here before on the fine print. We wrote about the lawsuit over what Claude's "5x" and "20x" plans actually promise, and the argument there was never the price — it was that nobody could tell what they had bought.
The generous reading, and the one that should worry SpaceXSI
Two explanations can be true at once: SpaceXSI throttled, and Cursor's dashboard lied about it. If the spikes were mostly a bug, Musk's reply is still the story, because it tells you exactly what happens the next time the meter moves: nothing, plus an explanation of the unit economics.
The contrarian case is that costs were always fine and this is pre-emption before a hard Q4 print. There is no way to check that from outside, because no cost figure and no revenue figure was given. That absence is itself a data point — a company confident in the ratio publishes it.
The honest summary is narrower than the outrage: a fast-growing agent product discovered that its heaviest users cost more than they pay, and its owner said so in public instead of quietly tightening the cap. The tightening is coming either way.
What to watch
Whether SpaceXSI publishes allowance sizes. It currently publishes the order of the tiers — 4x, 15x, 50x against Plus — and not one absolute number, which makes every complaint unfalsifiable in both directions.
The Q4 disclosure: paid weekly users against subscription revenue. If all 800,000 sat on the $30 tier, and none of them do, the line would be about $24 million a month before tax. That is the denominator the serving cost has to fit inside.
And the direction of travel across the labs. On November 1, Anthropic went the other way: Claude Max 20x became Ultra and the five-hour cap disappeared. One lab is shipping more allowance than it can meter, the other is admitting it can't afford the allowance it sold. Which one blinks at the next model launch is the pricing story of 2027, and Musk just moved first.
If your agent's weekly meter can jump 40 points overnight with no work done, what would you need to see before you'd trust it with a team's worth of tasks? Write to us via our contact page.
