Intelligence Per Dollar

Microsoft added a new metric called average token usage to its model release cards, presenting a new standard that considers both performance and cost simultaneously. For example, Microsoft's model achieves high performance on SWE-Bench Verified while using only about 1/3 the tokens compared to Claude Haiku 4.5.

AI benchmarks are now measured across two dimensions beyond simple performance measurement, including the cost of achieving a certain level of intelligence. This suggests that the era of reckless subsidies and tokenmaxxing—excessive token usage to boost benchmark scores—is coming to an end.

In fact, Uber exhausted its AI spending budget in just 4 months and restricted employees' AI usage, while Salesforce spent $300 million on Anthropic token costs and froze engineer hiring.

Ultimately, the core question for enterprise customers comes down to "What is the intelligence per dollar?" Model companies must compete on both performance and cost, and the application layer will ultimately compete on dollars per outcome, such as ticket resolution or code deployment.