Coinbase has transitioned its internal engineering infrastructure to run on Chinese open-weight AI models as the default, according to CEO Brian Armstrong. The crypto giant slashed its internal AI spending by nearly half while allowing developer token usage to continue growing at an exponential rate. Coinbase’s internal LLM gateway replaced hyper-expensive frontier systems with two Chinese models: GLM 5.2 (Zhipu AI), which scored 62.1 on the SWE-bench Pro coding benchmark and costs roughly $1.40 per million input tokens, and Kimi 2.7 (Moonshot AI), deployed for high-volume coding tasks. In comparison, Anthropic’s Claude Opus 4.8 charges up to $5 for the same volume. Armstrong noted that 91% of engineers never reached old usage ceilings. Coinbase re-engineered its gateway around three principles: intelligent routing, which reserves top-tier frontier models for complex tasks; aggressive caching, which increased the cache hit rate from 5% to 60%; and streamlined context engineering, requiring fresh chat sessions when switching tasks.