Enterprises are increasingly adopting model routing to assign AI tasks to the most cost-effective models to curb surging operational expenses.
This shift matters because it reduces the per-task revenue that frontier-model providers like OpenAI and Anthropic rely on to fund their operations and justify their planned initial public offerings.
Model routing allows companies to use expensive, high-capability models for complex reasoning while shifting routine tasks to cheaper alternatives. This trend has accelerated over the past two years [1, 2]. The need for such controls is underscored by extreme spending cases; one enterprise customer spent nearly $500 million on AI models in a single week, exceeding its spending limits [2].
For providers like OpenAI, the financial pressure is significant. The company does not expect to be profitable until at least 2030 [3]. While Anthropic recently hit a record valuation [2] and has reportedly surpassed OpenAI in enterprise AI spending for the first time [4], the rise of cheaper alternatives remains a threat.
Competition from abroad is intensifying this pressure. The Chinese AI lab DeepSeek provides frontier-level capabilities at a fraction of U.S. costs [5]. Some reports suggest Chinese models are taking a growing share of enterprise traffic, though others argue that routing primarily affects revenue rather than overall market share [5, 1].
Cognition CEO Scott Wu is among the AI enterprise leaders adopting these routing strategies to manage the missing return on investment associated with token billing [2], Wu said. As businesses prioritize the bottom line, the high-cost model of the early AI boom is facing a correction.
“Model routing is a fix for AI overspending that's a problem for OpenAI and Anthropic.”
The transition toward model routing signals a shift from the 'experimentation phase' of generative AI to an 'efficiency phase.' As enterprises demand a clear return on investment, the ability to swap expensive flagship models for cheaper, specialized ones erodes the pricing power of the industry's biggest players. This creates a valuation gap where the high costs of training frontier models may no longer align with the decreasing revenue generated from routine enterprise usage.
