Coinbase Halves AI Costs by Routing to Cheaper Models, No Engineer Caps

Coinbase cut AI spend 50% by defaulting engineers to open-weight models GLM 5.2 and Kimi 2.7, improving cache hit rate from 5% to 60%, and routing tasks by complexity. CEO Brian Armstrong shared five levers, noting 91% of engineers never hit old usage limits. The shift signals enterprise budget pressure driving adoption of cheaper models, threatening OpenAI and Anthropic revenue.