Coinbase Halves AI Costs: The Key Is Not Usage Friction but Infrastructure Optimization
On June 27, Coinbase CEO Brian Armstrong shared the company's core strategy for controlling AI spending. He argued that to keep AI costs stable while usage grows exponentially, the solution is not to introduce usage friction or spending alerts, but to optimize the underlying infrastructure: better default models, intelligent routing, and caching mechanisms. Through its self-developed LLM Gateway, Coinbase has reduced AI spending by nearly 50%, while token consumption continues to rise.
Default Model Migration and Routing: GLM 5.2 and Kimi 2.7 Take Center Stage
Coinbase's LLM Gateway now defaults to open-weight models such as GLM 5.2 and Kimi 2.7, while still allowing engineers to choose the best model per task. Data shows that 91% of employees never hit their usage limits, so instead of lowering quotas or adding warnings, the company shifted to cheaper default models. For routing, Coinbase preprocesses prompts in custom flows and routes tasks to the most cost-effective model based on cache hit rates and model pricing. For example, planning stages may require frontier models, but execution stages often do not. Armstrong believes humans should not manually select models in the future; AI can handle that automatically.
Cache Strategy and Context Pruning: Key Details for Cost Control
Armstrong highlighted that cache misses are the easiest way to drive up AI costs. All Coinbase requests are cache-aware to maximize reuse of hot caches. For instance, after implementing proper caching for LibreChat, the cache hit rate jumped from 5% to 60%. Additionally, Coinbase requires engineers to keep context lean, including starting new sessions when switching tasks, narrowing file context scopes, and disconnecting unused tools. The goal is not to suppress AI usage but to build infrastructure that can support exponential growth. Through these measures, Coinbase has cut AI spending by nearly half while token usage continues to grow, validating the effectiveness of the approach.