The CEO of Coinbase Brian Armstrong disclosed on X that the company had set Kimi 2.7 on the dark side of GLM 5.2 and the moon as a default model for engineers. It states that AI expenditure has been reduced by almost half, and token usage continues to grow, in combination with route optimization and cache improvement measures.
Default Model Switch
Coinbase states that these two models are used mainly for routine tasks. For more complex planning categories, engineers can still choose a more advanced model. Armstrong believes that many implementation scenarios do not require the most expensive model as a default option.
Three reduction measures
The practices listed include smart routing, active cache and streamlined context. The system automatically distributes tasks on the basis of a hint, a cache-rate and model pricing. Coinbase also requires that the request have a cache sensory capacity and re-use the existing cache to the extent possible.
The code review uses multiple models.
In the code review segment, Coinbase uses multi-model parallel validation results to maintain quality standards. Armstrong also stated that the company would visualize the use data and link the use to business impact, emphasizing that the goal was to expand the use of AI rather than to compress the use itself.