AI costs explained: why 'tokenmaxxing' is ending and cheaper models are winning

For the past two years, the unofficial motto inside a lot of AI-hungry companies was simple: use as much of it as you possibly can. Staff were nudged, and sometimes openly rewarded, for running every task through the priciest, most powerful model going. There was even a nickname for it — 'tokenmaxxing'. In late June 2026, CNBC reported that the mood has flipped. The same firms that were egging everyone on are now quietly counting the cost, and some are switching to far cheaper models made in China. Here's what's happening, in plain English, and why it actually matters to you.

The trouble with pay-per-token is that costs can run away from you. If your team leans on a top-tier model for everything — drafting emails, summarising documents, writing code, running automated 'agents' that loop over a task dozens of times — the meter never stops. And because the labs behind the leading American models, OpenAI and Anthropic, have been slower to cut their prices lately, that meter kept spinning at full rate. CNBC reported that big names including Uber, Microsoft, Salesforce and Meta have started rationing how much advanced AI their employees can use, because the pay-per-token model turned out to be 'more expensive than it's worth'.

Uber's story is the one that makes the point stick. Its chief technology officer, Praveen Neppalli Naga, said the company blew its entire annual AI budget in just four months, according to The Information. Uber's fix was to bring in spending tiers — a baseline of around $1,500 a month per employee, with more available if you can justify it. Think of it like a company mobile plan: you get a generous allowance, and if you want to go over it, you have to explain why. It isn't banning AI; it's asking the question that got lost in all the excitement — is this spend actually paying for itself?

The other half of the story is where the money is going instead. A growing number of companies are switching to 'open-weight' models — and many of the best-value ones right now come from China. An open-weight model is one whose 'brain' (the trained parameters, or weights) you can actually download. Instead of renting access through someone else's cloud and paying per token forever, you can run the model on your own servers. That matters for two reasons: it can be dramatically cheaper at scale, and your data never has to leave your building.

Take Lindy, a roughly 25-person startup that builds AI agents. Its chief executive, Flo Crivello, moved 100% of the company's traffic off Anthropic's Claude and onto DeepSeek, a cheaper Chinese open-weight model. He told CNBC the cost curve 'crash[ed] to the ground', said he expects to save millions within months, and bluntly called the switch 'a matter of survival'. One analysis put the saving on the migrated work at around 90%. He'd happily switch back if American prices fall — 'until then, we've got options'. And the scale of the move shows up plainly in figures from OpenRouter, a popular service that routes app requests to whichever model you pick:

In plain terms: of all the tokens flowing through the ten most-used models on that platform, the Chinese-built ones — Qwen (Alibaba), DeepSeek, Kimi (Moonshot), GLM (Zhipu) and MiniMax — now account for roughly 61%, about six in every ten. Under two years ago it was a rounding error.

If you use AI tools day to day — or you're the one paying for them — this shift is mostly good news. More competition between models means downward pressure on prices, and you're already seeing it: Anthropic's new Claude Sonnet 5 launched at a cut-price $2/$10 per million tokens, and OpenAI previewed a cheaper 'Terra' tier, both aimed squarely at this pressure. More choice means you're not locked into a single provider. And the return of the ROI question — does it actually pay off? — is healthy: it pushes tools to earn their keep instead of dazzling you with horsepower you don't need.

There's a sensible flip side worth keeping in mind, though. Cheaper isn't automatically better for every job — a bargain model that gets things subtly wrong can cost you more than it saves. Running an open-weight model yourself takes real engineering muscle, so it's not a switch most individuals or small teams will flip themselves. And the tug-of-war between American and Chinese labs is tangled up with export controls and politics, so today's cheapest option might not stay the cheapest. The practical takeaway is refreshingly old-fashioned: pick the tool that does your job well for the least money, check now and then that it's still the best deal, and don't pay for horsepower you're not using.