For years, AI firms gave customers unfettered entry to the sweet retailer, encouraging them to think about tokens, the chunks of textual content AI reads and writes, as successfully infinite.
Tokens had been bundled into subscriptions, hidden behind beneficiant caps, or priced low sufficient that individuals stopped counting them. However as the price of serving fashions eats into income, and as chip shortages, helium disruption, and knowledge middle bottlenecks constrain how a lot compute can come on-line, the massive mannequin makers are beginning to ration entry extra aggressively. All-you-can-eat AI is disappearing. Now firms are in a contest to see who can preserve subsidising demand the longest, and whether or not the final to blink will get to dominate the market.
This week, Meta took offline its “Claudenomics” leaderboard, which tracked worker productivity utilizing a crude metric of what number of AI tokens they used over the previous month. Staff used greater than 60 trillion tokens in a single month, equal to round 80 million copies of Warfare and Peace, or the contents of 10,000 total libraries.
“Main frontier mannequin builders are going to face trade-offs in how they use their compute sources,” explains Sam Manning, senior analysis fellow at GovAI, a group of researchers finding out how AI is used and deployed. “It’s an excellent consequential determination these firms have to make.”
The worldwide scarcity of AI chips, more likely to be exacerbated by the Center East conflict’s impression on helium, a key part in GPU manufacturing, together with a backlog in constructing knowledge facilities, means there’s solely a finite quantity of {hardware} to each practice and run AI fashions. Dial down the coaching price range and also you threat falling behind opponents in releasing cutting-edge fashions. Reduce on inference, the velocity and scale at which you meet buyer demand, and also you frustrate customers.
Completely different firms are taking completely different approaches. Earlier this month, OpenAI introduced it would switch users on its Codex app to token-based pricing, fairly than per message, no matter question measurement. That might profit these working smaller duties, however might additionally shortly burn by means of a consumer’s token allowance. The corporate additionally ended a months-long supply to double Codex limits initially of April.
Across the similar time, Anthropic blocked users from using Claude subscriptions to energy OpenClaw agentic AI instruments, pushing them as a substitute towards API entry. The possible cause was easy: demand. “We’ve been working exhausting to fulfill the rise in demand for Claude, and our subscriptions weren’t constructed for the utilization patterns of those third-party instruments,” mentioned Boris Cherny, Claude Code govt, asserting the shift. “Capability is a useful resource we handle thoughtfully and we’re prioritizing our clients utilizing our merchandise and API.”
The monetary strain is evident. The price of serving AI fashions accounts for greater than half of OpenAI and Anthropic’s revenues, in response to inside knowledge obtained by the Wall Avenue Journal. “There’s simply been enormous client surplus,” says Manning. “Quite a lot of the preliminary motivation for pricing was to construct up market share and get customers onto their platforms. Perhaps it’s the case that we’re seeing some kind of an inflection level there.”
The worth-versus-performance trade-off is just not restricted to U.S. corporations. It is usually entrance of thoughts for China’s AI firms. Zhipu AI, which makes the GLM fashions, has seen its open platform API token costs rise 83% year-to-date in early 2026, and this week introduced another 8% increase for its newest fashions.
The worth hikes replicate accelerating demand, in response to JP Morgan analysis. Customers seem prepared to soak up larger prices for higher-value workloads, significantly in coding and agent-related use circumstances. Rising costs and sustained demand are already reshaping unit economics for China’s AI giants, with Zhipu AI’s API gross margins increasing from 3% in 2024 to 19% in 2025.
Nonetheless, Alibaba is taking a unique tack. The corporate has made its Qwen-3.6 mannequin free to customers by means of OpenRouter, a coding help system. Customers shortly burned by means of nearly 1.5 trillion tokens in a single day.
That call stands out, however the logic is evident. Alibaba is attempting to win builders, workloads, and long-term cloud clients. Whereas OpenAI and Anthropic are tightening entry to guard scarce capability and enhance unit economics, Alibaba is taking part in an extended recreation, absorbing the price in hopes of locking in customers which may be more durable to win later.
Alibaba might additionally profit from the very fact most firms can’t compromise on worth any time quickly—if ever. Pricing pressures stay unavoidable if compute stays scarce, in response to GovAI’s Manning. “We should always anticipate there to be this kind of shortage of compute for the foreseeable future,” he says.
