Coinbase CEO: Controlling AI Costs by Optimizing Default Models and Routing Mechanisms to Achieve Growth in Token Usage and Reduction in Spending
Odaily Odaily reports that Coinbase CEO Brian Armstrong posted on platform X, stating that within Coinbase, teams are exploring how to keep AI spending stable while token usage grows exponentially. This is achieved not through usage limits or frequent cost reminders, but through better system design.
He pointed out that Coinbase optimizes AI usage costs in five main areas: First, "better default models," allowing engineers to freely choose models, but the system defaults to guiding them towards lower-cost open-source weight models (such as GLM 5.2, Kimi 2.7), while still encouraging selection of appropriate models for different tasks; Second, "smart routing," using an automated system to allocate the optimal model based on task type, cache hit rate, and model price, rather than manual selection.
Third is "cache optimization," significantly reducing repetitive computation costs by improving the cache hit rate; Fourth is "streamlined context," reducing wasteful token consumption during task switching and limiting unnecessary tool calls; Fifth is "visual transparency," allowing engineers to clearly see their own usage, but without suppressing usage through restrictions.
Brian Armstrong stated that the goal is not to suppress AI usage, but to build infrastructure capable of supporting exponential growth. Results show that with continuous growth in token usage, overall AI spending has nearly been halved.
