Uber burned its entire 2026 AI budget in four months. The same math hits small teams sooner, with far less cushion. Here’s how to price an agent’s compute before it prices you.
Carter Busse watched his company’s AI bill jump 7x in a single day. Workato, the software company where he’s CIO, had handed AI agents to its staff the previous summer, and adoption spread fast, with people rebuilding their own jobs around them. Then in May, Anthropic moved the company onto token-based pricing, where you pay for what you consume instead of a flat seat fee. Spend went up sevenfold overnight. “We created a monster,” Busse told the Financial Times.
That story is playing out everywhere, and the names attached to it are not small. Uber gave roughly 5,000 engineers an AI coding tool late last year, watched adoption rocket past 80%, and by April had spent its entire 2026 AI budget. The whole year, gone in four months. It now caps employees at $1,500 a month per tool. The Financial Times reported in June that Amazon, Walmart, Cisco, Uber, and Meta are all clamping down at once, with Meta’s CTO telling staff that piling up tokens is no measure of impact. A year ago, Nvidia’s Jensen Huang said he’d be “deeply alarmed” if a $500,000 engineer wasn’t burning $250,000 a year on tokens. The mood turned that fast.
Why this lands on you first
The part that should bother you more than the Uber headline: the industry sold agents to small teams as the great equalizer, the thing that finally lets a five-person marketing team punch like a fifty-person one. And that pitch is half true. But the cost lands hardest exactly where the budget is thinnest, and it lands first on the people with no buffer to catch it.
A big company that blows its budget in four months has a finance team, a board, and a quarter of runway to react. You don’t. When you wire up an agent that runs every time a lead comes in, or rewrites every product description nightly, the meter starts spinning against a number you feel. And if you’re the operations person inside a growing company, building the clever agent everyone’s excited about, the token bill is about to become the thing finance watches. Which means the crackdown lands on you, as a governance problem, right when you thought you’d shipped a win.
I know the shape of this because I ran straight into it. I have an agent I call Metis: it lives on a machine in my apartment and works through tasks on its own, reading, acting, checking its work, trying again. In its first three weeks, with every task routed through Claude’s API, it burned more than $300. Not because anything broke. Because that’s what an agent that thinks in a loop costs when you let it run. At its peak it was hitting $30 a day, on a path to roughly $900 a month, and the single biggest offender was a newsletter job that processed everything in one giant call. Nobody warned me at the design stage. The number just showed up on a dashboard. If you’re building something similar, it’ll surface on yours the same way: quietly, and after you’ve already shipped it.
This isn’t a “spend smarter” problem
The advice rushing in to meet this moment is mostly cost hygiene: send cheaper queries to cheaper models, cache repeated calls, batch your requests. It works. Teams report cutting bills by half or more that way, so do it. But notice what it is: a way to shrink a bill after you’ve already committed to the thing generating it. It treats the symptom.
The deeper miss is that almost nobody prices an agent’s compute before building it. We still think about AI the way we thought about software, as a predictable monthly seat. An agent isn’t a seat. A chatbot takes your question and answers it once. An agent takes your goal, decides its own sub-tasks, runs them through a chain of model calls and tools, and only stops when it’s done, and every step in that chain is a separate charge. The price of a token keeps falling. The number of tokens one task burns keeps climbing. That gap is where the surprise bills live, and at the extreme it’s how one company reportedly ran up a $500 million Claude bill in a single month after nobody set a cap (per Axios, sourced to a single anonymous consultant, so treat it as folklore). The point isn’t the number. It’s that an uncapped loop has no natural ceiling.
Price the agent before you build it
So price it at the design stage, the way you’d price a hire. Three questions, in order.
What does one run cost? Don’t try to estimate this from theory; almost nobody can guess token consumption in advance, and you don’t have to. Run the agent a handful of times, read the actual spend off your provider’s dashboard, then multiply by how often it’ll fire in production. An agent triggered on every inbound lead at a company doing two thousand leads a month is a very different line item than one you run by hand on Fridays. You don’t need a precise figure, just the order of magnitude: coffee money, or a part-time salary? For a sense of how wide the range runs, my own agent swung from $30 a day at its busiest down to near zero for the work I later moved off the paid API. Same agent, same kinds of task. The difference was entirely in how hard it looped and which model it looped on.
Is the run worth more than it costs? This is the only question that actually decides anything, and there’s no universal number I can hand you. The threshold is yours to work out. Put the monthly compute cost next to the value of what the agent produces: the hours it gives back at your loaded labor rate, the revenue it touches, the work you’d otherwise pay someone to do. One of my agent’s research runs cost about $4 to $5 in API calls and produced a prospect brief I’d have paid a person far more to assemble. Easy yes. That same $5 spent to shave four minutes off a task I do twice a week? Easy no. Most people never run this comparison, because the cost was invisible until the invoice.
If it doesn’t clear, what’s the move? You have more than one.
- Drop to a smaller model for the sub-tasks that don’t need a frontier brain. Most of an agent’s steps are clerical, and a fast, cheap model handles them fine. The discipline is matching the model to the job, not routing everything to the biggest one out of habit.
- Redesign the loop so it makes fewer calls: cache, batch, stop it re-reading the same context ten times. When my newsletter job was the budget hog, the fix was breaking one giant call into smaller pieces. This is the hygiene layer, and it belongs here as one option among several, not as the whole answer.
- Kill the agent. The option nobody writes down. Sometimes the honest call is that the task wasn’t worth automating, and a saved prompt you run by hand does the job for a fraction of the cost.
- Own the compute. At the small end, you can run capable open models on your own hardware and take the per-token meter off the table for the work that doesn’t need a frontier model. After that $300 lesson, I moved Metis off a small machine that couldn’t run anything local and onto one with enough memory to run open models well. The morning digests and routine research now run locally at no per-task cost; only the work where quality matters still hits the paid API. My monthly spend went from a $900 path to around $25. The honest catch: the hardware wasn’t free, and it paid back over months, not on day one. For a thirty-person company this usually isn’t today’s move. But it’s worth knowing the ceiling is escapable, because the whole industry above you is discovering it can’t easily escape.
The dependency you haven’t broken yet
If you’ve already done the work of unbundling your stack, picking each tool for its job instead of swallowing whatever came in the suite, you know the value of not being locked in. The token bill is the next chapter of that same story. You broke free of the bundle. The metered API is the dependency you haven’t looked at yet, and it’s the one quietly setting your variable costs.
You don’t have to self-host tomorrow to act on this. You have to do the one thing every company in the headlines failed to do: put a price on the agent before you build it, and check that the price is worth paying. Big tech skipped that step with a year of runway to absorb the mistake. You won’t have the runway. You do, though, have the advantage they’re all suddenly jealous of. You can see the meter, and you’re small enough to do something about it before the bill arrives.
I run the AI Strategist newsletter on the same principle: what the headlines miss, priced for people who actually have to make the call. If you’d rather catch the next cost trap before it shows up on your dashboard, that’s where I write it down first.
Sources and Further Reading
- Rebecca Bellan, “The token bill comes due: Inside the industry scramble to manage AI’s runaway costs,” TechCrunch, June 5, 2026
- Nora Young, “Their goal was: use as much AI as possible. Now some companies are pulling back,” CBC News, June 17, 2026
- “The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI,” 404 Media, June 2026
- Jamie John, Rafe Rosner-Uddin, and Ryan McMorrow, “‘We created a monster’: companies rein in AI usage as costs strain budgets,” Financial Times, June 19, 2026.
- “AI sticker shock hits corporate America,” Axios, May 28, 2026
