Coinbase is nailing cost reduction and efficiency: slicing open-source models, implementing smart routing, and boosting cache hit rates from 5% to 60% — a textbook case for AI infrastructure.

According to BiJie.com news, Coinbase CEO Brian Armstrong said in a post on X that Coinbase has reduced AI spending by nearly 50% while token usage continues to grow by optimizing default settings, routing, and caching strategies. Specific measures include: switching the default model to open-weight models such as GLM 5.2 and Kimi 2.7; 91% of employees previously had never reached the usage limit, but in a custom system they pre-process prompts and automatically route them to the most suitable model; implementing differentiated handling for planning and execution tasks to improve cache hit rate; increasing LibreChat’s cache hit rate from 5% to 60% by streamlining context management; when switching tasks, opening a new session and narrowing the file context scope to enhance spending visibility; engineers can freely choose models but must bear the corresponding expected impact.