Businesses turn to cheaper AI models as usage bills rise

OpenAI was considering drastic price cuts on June 10 as competition from Anthropic sharpened, a sign that the AI pricing fight had moved from model performance to enterprise budgets. Businesses are now steering routine work toward cheaper, smaller and more specialized systems as usage bills rise, and the question inside procurement teams has shifted from which model is most powerful to which one is good enough at a sustainable price.

That change is visible in the way corporate buyers are using AI. Microsoft’s Satya Nadella, Palo Alto Networks’ Nikesh Arora and Coinbase’s Brian Armstrong have each argued that leaner models can handle a large share of everyday tasks, including customer support, coding assistance and document review. During the first wave of AI enthusiasm, some companies treated heavy token consumption as a sign of productivity and chased tokenmaxxing. Now the bill is showing up in monthly spend, and the economics are forcing a reset.

The pressure comes partly from the pricing structure itself. Token prices have been falling, but the total cost of finishing tasks has not followed the same path because more providers are moving from flat subscriptions to usage-based pricing. That leaves enterprise budgets less predictable and pushes buyers toward routing tools that send simple requests to cheaper models while reserving premium systems for harder work. In practice, that means more scrutiny from finance teams that are being told to prove return on investment and to rein in out-of-control token spend.

The shift is showing up in market data. A Citi note cited in the coverage said open-source models processed on OpenRouter accounted for 65% of tokens in June 2026, up from 34% in January. That jump points to a faster move toward cheaper alternatives, including open-source systems such as DeepSeek, which are gaining traction on cost even as security concerns still limit adoption at some larger businesses.

For frontier AI firms, the change weakens the premium-model moat they spent two years building. If routine corporate work can be handled by cheaper systems, the remaining pitch for expensive models has to rest on clear performance gains, not just scale. The pressure now extends beyond model makers to cloud providers, app developers and cybersecurity teams that built their forecasts on rapid AI adoption.