Back to News
CNBC

Smart spending on AI: How 'model routing' is cutting costs and reshaping the market

Corporate America is waking up to a hard truth about artificial intelligence: the bills are piling up faster than anyone expected. CFOs and boards, once content to let teams default to the most...

Corporate America is waking up to a hard truth about artificial intelligence: the bills are piling up faster than anyone expected. CFOs and boards, once content to let teams default to the most powerful AI models for every task, are now demanding a more disciplined approach. The shift is already starting to reshape the AI industry.

The fix gaining traction is called model routing. Instead of sending every query through a premium frontier model—whether it's a complex coding problem or a simple request like naming the third U.S. president—routing tools match each job to the most cost-effective model. Easy tasks go to cheaper, faster alternatives; hard problems still get the expensive firepower.

Scott Wu, CEO of Cognition, says the savings on routine work are dramatic. For boilerplate tasks, companies can cut costs five to ten times by using models that are good enough for the job. Yet most enterprises aren't routing at all. Glean CEO Arvind Jain estimates that 95% of enterprise AI usage still runs on the most costly models.

The pressure is real. Cisco's chief product officer Jeetu Patel broke down the math: at roughly $200 per employee per week in token usage, a 90,000-person company faces a $900 million annual bill. Cisco came in well over budget and has had to reallocate resources, prioritizing tokens over other spending.

This shift spells trouble for AI leaders like OpenAI and Anthropic, which have built their business models on premium pricing for every query. If companies steer high-volume, simple work to cheaper open-source models, the frontier labs lose a huge revenue stream. Patel predicts the pricing model will have to change, with labs getting more efficient rather than simply charging more.

Cognition is already responding with an unusual guarantee: if its Devin agent doesn't deliver engineering value equal to what customers pay, Cognition will fund usage up to $10 million to make up the difference. The message is clear: companies should pay for output, not activity.

The question now is how much of the market is routine work versus complex problems. The answer will go a long way toward determining the valuations of the leading AI companies. Pricing power is moving from sellers to buyers.