In June 2026, the fintech startup Slash - valued at $1.4 billion - encouraged its team to use AI coding tools more. Nicolas Brillante, its head of strategic verticals, took that to heart and spent a day building a playable meme game called "Brainrot Shooter" with Claude.
By the end of the week, that experiment had run up $81,267 in Claude tokens. Brillante posted the bill publicly, calling it a genuine accident: he had underestimated how fast the cost adds up when you repeatedly load a full codebase into an AI agent during active development. The story went viral.
Why one week cost $81,000
A chat costs a few cents. An AI coding agent is a different animal. It reads your files, writes code, runs it, reads the errors, and tries again - and every step re-sends the growing pile of context back to the model. A long session on a big codebase re-bills the same files over and over.
The cost curve is the trap. A session's tenth step costs more than its first, because the context has grown. The hundredth costs more still. Human intuition prices work linearly - twice the work, twice the cost. Agent billing is closer to a snowball, and the snowball keeps rolling while you get coffee.
Do that all day, for a week, and the meter you never see spins into five figures. Brillante is a smart engineer at a billion-dollar company. That is exactly the point: this is not a skill problem, it is a visibility problem.
Scale it down to a normal company
You may never see an $81,000 week. The quieter version of the same problem is probably running in your company right now. Take a 10-person dev team using AI coding tools the way Microsoft's engineers were - in the $500-2,000 a month range each - and do the arithmetic:
- Monthly team spend: $5,000 to $20,000.
- Over a year: $60,000 to $240,000, metered out token by token with nobody watching any single day of it.
- The waste share: industry research finds 30-40% of token spend buys nothing extra. On these numbers, that is $1,500 to $8,000 a month leaving quietly.
The waste is not mysterious. Whole codebases loaded into context when one file would do. Premium models on boilerplate, when the per-token price gap between model sizes runs 10-30x. The same instructions re-billed on every agent step, which caching can cut by up to 90% on repeated prompts. None of it makes an $81k headline. All of it compounds the same way.
New to how token billing works in the first place? The plain-English tokens guide covers it in five minutes.
It is not just startups
If you assume the giants have this figured out, they do not. The biggest, most technical companies in the world got caught by the same meter this year:
- Microsoft cut most internal Claude Code licenses across its Experiences & Devices group (Windows, Microsoft 365, Teams, Surface) with a June 30 deadline, after token costs blew through the budget - some engineers were running $500-$2,000 a month.
- Amazon - one of Anthropic's own biggest investors - now requires explicit sign-off before engineers use Claude Code on production work, steering them to a cheaper in-house tool. It also scrapped an internal leaderboard that ranked staff by AI usage once people started gaming it and driving compute costs up.
- Uber burned its entire 2026 AI budget in four months and now caps engineers at $1,500 a month per tool.
Startup or trillion-dollar enterprise, the failure is identical: thousands of people (or just one) with unmonitored, uncapped access to a tool that bills by the token.
"Our engineers are careful" - the objection that doesn't survive this story
Every company on this page would have said the same thing, and meant it. Brillante is a capable engineer at a $1.4 billion company. Microsoft and Amazon employ some of the most careful engineering organizations on earth. All of them got surprised by the same meter.
Careful people still cannot see an invisible number. An agent gives no feedback while it spends - no progress bar, no running total, nothing on screen that says what the session has cost so far. The tool feels exactly the same during a few-cents chat and an $81,000 week. Care without instruments is just hope.
So the fix is not better people, and it is not a lecture about prompt discipline. It is a meter someone can actually read - which is why the first thing to change is when the number becomes visible, not who is allowed to spend it.
What actually prevents this
Not banning AI - every company on this list wanted the productivity. What was missing was the boring infrastructure around it:
- Live visibility - spend per person and per project, updated daily, so an $80k week shows up on day one, not on the invoice.
- Alerts - a ping when someone's run rate goes vertical, while you can still do something about it.
- Caps - sensible per-person limits that turn a $81,000 surprise into a $2,000 heads-up.
How to check your own exposure this week
Before writing any policy, answer one question: if someone here had an $81,000 week, when would you find out? Five checks get you to an honest answer:
- Pull last month's usage by person. Every major AI provider has an admin console or a usage export. Sort it by spend. The shape of the list matters more than the total.
- Find your spread. Compare your top spender to your median one. A big gap is normal once agents are involved - an invisible big gap is the risk.
- Time the lag. Count the days between tokens being spent and a human seeing the number. For most companies the honest answer is "when the invoice arrives" - up to a month of blind spot. Slash's lag was one week, and it cost $81,267.
- Ask who runs agents. Long sessions, parallel agents, overnight runs - the same habits that built Brainrot Shooter. The goal is to know, not to forbid.
- Check the alert path. If someone's spend went vertical tomorrow, does anything ping anyone? If the answer is no, you have the same setup Slash had, minus the viral post.
None of this requires a project. Connecting spend data takes about 15 minutes, read-only, with the first full report inside 48 hours and the first real cut typically about two weeks in. That is the boring infrastructure - the kind covered in our AI cost management overview - that turns the next $81,000 week into a same-day alert. Every company in this story built it after the bill. The cheaper order is before.