"Tokenmaxxing Lost Its Cool"

There was a brief, strange moment in enterprise AI where burning through tokens became a point of pride. Engineers at some companies reportedly competed over who could rack up the biggest model bills, and "tokenmaxxing" — the practice of hitting frontier models so hard your usage graphs looked vertical — was a badge of engineering ambition. That moment, according to a sharp piece in The Economist this month, is over.

Uber dropped the mic on this particular party: the company disclosed in April that it had already spent its entire annual AI budget in four months. That's not a rounding error. That's a CFO walking into a meeting with a spreadsheet and saying, in effect, we need to have a conversation about pacing. Uber's not alone. One unnamed firm reportedly burned through $500 million on AI, and others are experiencing variations on the same theme — budgets that were supposed to last a year evaporating in a quarter.

The culprit isn't just more usage; it's structurally more expensive usage. Reasoning models and AI agents don't behave like old chatbot queries. They chain together dozens or even hundreds of model calls to complete a single task, and in some cases agents are building their own agents, creating a recursive cost amplifier nobody fully priced in. The sticker shock has been sharp enough that companies like Meta and Amazon have quietly scrapped internal leaderboards — not because they stopped caring about model performance, but because optimizing for raw capability without a cost lens started looking reckless.

What comes next is already visible. Software vendors that bake AI into their products are experimenting with outcome-based pricing instead of raw metered billing, trying to reassure clients that the bill won't spiral unpredictably. More companies are getting selective about model choice, realizing plenty of tasks don't need a frontier reasoning model when a smaller, cheaper one works fine. And hanging over all of this is the expected public-market debut of Anthropic and OpenAI later this year — both of which will face pressure to actually turn a profit, which means one thing for their customers: prices are going up, not down, exactly when everyone is scrambling to cut back.

The vibe shift is real. A year ago, the question was "how fast can we adopt AI?" Now it's "how much is this actually costing us, and what do we turn off first?" Tokenmaxxing, it turns out, was a lot more fun when someone else was paying the bill.

Further reading: The Economist has the full story, and the Mint syndication covers the agent-cost spiral in detail.