Soaring AI bills push businesses towards cheaper models

TL;DR:

• Tech leaders including Microsoft’s Satya Nadella now argue that smaller, cheaper models can handle much of what businesses need.
• A shift from flat subscriptions to usage-based pricing is leaving firms with unpredictable, often higher bills — Uber reportedly spent its entire 2026 AI budget in four months.
• Open-source tokens processed on the OpenRouter marketplace jumped to 65% in June, from 34% in January, as companies route cheap work away from premium models.

The reflex to buy the most powerful AI model available is fading. A growing number of technology executives now argue that cheaper options are what will actually drive adoption, after a year in which heavy consumption — sometimes treated as a proxy for productivity, and nicknamed “tokenmaxxing” — collided with the bills it generated.

When the meter starts to bite

The headline price of tokens, the units used to measure AI usage, keeps falling. But the cost of completing a task is rising, because providers are moving from flat subscriptions to usage-based pricing and because each task now involves more steps, data and longer inputs. The result is bills that are hard to forecast: Uber burned through its entire 2026 AI budget in four months after staff rushed to adopt coding tools, forcing managers to cap usage, according to reports cited by Reuters. Gartner estimates AI coding costs will overtake the average developer’s salary by 2028.

That pressure is pushing firms towards routing tools such as OpenRouter, which assign cheap work to the most cost-effective system and reserve premium models for complex jobs like coding. The four most popular models on the marketplace are now Chinese, led by DeepSeek, with open-source providers charging as little as 18 cents per million tokens against roughly $4 for top US models. The capability gap, one executive noted, has narrowed from more than a year to about four months.

For UK businesses, the lesson lands the same way as the AI-spending jitters that hit tech stocks last week: the question has shifted from whether to use AI to what each task is actually worth. Procurement teams weighing model contracts now have to model consumption, not just licences — and weigh security concerns that still make many wary of Chinese open-source options.

Looking forward

A price war looks likely as OpenAI and Anthropic, both eyeing public listings, compete on cost as well as capability — OpenAI has reportedly weighed significant cuts. Expect more firms to follow the cloud-computing playbook: spreading work across multiple providers in search of the best fit and price, rather than betting on a single frontier model.