Open-source model + intelligent routing + caching as the three-pronged approach, Coinbase's AI costs are directly halved. Brian Armstrong's sharing is worth copying.

Coin World News, Coinbase CEO Brian Armstrong shared practical experience in maintaining stable AI spending against the backdrop of exponential growth in token usage. He stated that the company did not adopt restrictive measures such as setting usage caps, but instead achieved cost reduction and efficiency gains by optimizing default models, smart routing, and caching strategies: In terms of default models, open-weight models such as GLM 5.2 and Kimi 2.7 were introduced to replace expensive general-purpose models; routing mechanisms match models based on task requirements; cache pre-processing and task session management reduce token waste (cache hit rate increased from 5% to 60%). Thanks to these optimizations, Coinbase's AI spending has nearly halved, while token usage continues to grow.