No, a company didn't spend $500M on Claude in a month

A company spent $500 million on Claude in a single month, by accident, because nobody set a usage limit. That was the story last week, and it was everywhere. Reddit threads, X, and a long line of news sites all ran it inside a few days, each one citing the one above it.

It traveled that fast because people wanted it to be true. Every board that signed off on an AI budget last year is now asking what it bought, and plenty of pilots have little to show. A half-billion-dollar accident is the proof a skeptic has been waiting for, handed over gift-wrapped. It confirmed what they already suspected, so almost nobody stopped to check it.

Follow it back and there is almost nothing there. One AI consultant, unnamed. One client, also unnamed. One sentence, in a newsletter. No invoice, no token count, no company anyone can call. By the time it reached the headlines it had passed through several hands, and not one of the outlets running it was any closer to the source than you are.

The number is also impossible. Not improbable. Impossible for a single company in a single month.

The math doesn’t hold

Take Anthropic’s current rate for Opus 4.8: $25 per million output tokens, the most expensive line on its price sheet. Bill every token at that rate, all output, nothing cached, nothing discounted, the most generous assumption you can make for the claim. Even then, $500 million buys 20 trillion tokens in a single month.

Spread 20 trillion tokens across thirty days and it comes to 7.7 million tokens every second, without pause, all month. A fast Claude session puts out somewhere around sixty tokens a second. To reach the number you would need more than a hundred thousand sessions running flat out, none of them ever stopping, for thirty straight days.

Put it in people instead. Say one engineer leans on Claude hard enough to run up $2,000 a month, already a steep bill for a single person. To reach $500 million you need 250,000 of them, every one maxed out, every working day, for a month. A quarter of a million people doing almost nothing but feeding Claude all day. No real company looks anything like that.

What’s actually real

The sticker shock underneath the story is real, and it comes with names. Microsoft canceled most of its Claude Code licenses, partly over cost, according to The Verge. Uber’s COO said the company’s AI costs are getting harder to justify. Both name a company and an executive, and both describe bills in the thousands per engineer, not the millions per company.

That is the real shape of the problem: a number that climbs fast and still stops well short of nine figures. A team of engineers each running $500 to $2,000 a month adds up to a figure a CFO notices. It does not add up to half a billion dollars, and no amount of forgotten usage limits gets it there.

Why a fake number travels

The half-billion figure spread because it flatters two crowds at once. The skeptics get to say the bubble is real and someone just proved it. Everyone else gets to feel safe, because nobody runs up $500 million by accident, so the story becomes a freak event at a company nothing like yours. Either reading is an excuse to stop watching your own numbers, which is the one place this actually matters.

The overrun that should worry you is two orders of magnitude smaller and far more ordinary. It comes from poor context management and a plan no CTO pressure-tested before it reached production, not from a forgotten switch. It doesn’t announce itself at half a billion. It shows up at fifty thousand, then two hundred thousand, under the dashboards that were supposed to catch it, because the one control nobody set was a ceiling the agent has to respect while it runs.

No company spent $500 million on Claude last month. When a runaway AI bill finally gets a real name attached, it will have four zeroes, not nine. That one won’t trend or land a headline on a dozen sites. It is already sitting on someone’s invoice right now.