Coinbase has cut AI spending nearly in half and is trying to set open-weight models like GLM 5.2 and Kimi 2.7 as default options.

BlockBeats News, June 27 – Coinbase CEO Brian Armstrong stated that if the goal is to keep AI spending stable while token usage grows exponentially, the key is not to add friction or spending reminders, but to adopt better default models, routing, and caching mechanisms. Coinbase is experimenting with using open-weight models such as GLM 5.2 and Kimi 2.7 by default through its LLM gateway, while still encouraging engineers to choose the appropriate model based on the task. He noted that 91% of employees have never hit the usage cap, so instead of lowering limits and increasing reminders, the company has shifted to lower-cost default models.

In terms of model routing, Coinbase preprocesses prompts in custom workflows and routes tasks to the most suitable model based on cache hit rates and model pricing. For example, the planning phase may require a frontier model, but using a frontier model during the execution phase could be excessive. Armstrong believes that in the future, humans should not be the ones choosing models—AI can automate that task.

Armstrong also stated that cache misses are the easiest way to drive up costs. All Coinbase requests are cache-aware to maximize reuse of hot caches. For instance, after properly implementing caching, LibreChat's cache hit rate increased from 5% to 60%. Additionally, Coinbase requires engineers to keep context lean, including starting new sessions when switching tasks, narrowing file context, and disconnecting unused tools. The goal is not to suppress AI usage, but to build infrastructure that supports exponential growth. Through these practices, Coinbase has nearly halved its AI spending while token usage continues to grow.