Coinbase CEO Brian Armstrong says the company is moving more AI workloads to Chinese models including GLM 5.2 and Kimi 2.7. An automated router chooses models by task and price, while caching improvements reportedly raised the hit rate from 5 percent to 60 percent.
The company says these changes have cut AI spending in half. That is notable because many AI buyers are now comparing frontier-model quality against rapidly improving lower-cost alternatives.
For model providers, the Coinbase example points to a tougher pricing environment. Enterprise customers increasingly have the tooling to route around expensive defaults when cheaper models are good enough.