Coinbase CEO Brian Armstrong has revealed the mechanism that allowed the company to nearly halve its artificial intelligence costs, despite exponential growth in compute token consumption. According to him, the key to success lies not in strict limits and bureaucratic bans, but in smart architecture: default settings, request routing, and aggressive caching.
Armstrong emphasized that Coinbase engineers can choose any model for their work, but default configurations are crucial. The company actively experiments by setting lightweight models with open weights, such as GLM 5.2 and Kimi 2.7, available through an internal gateway, as the default. Surprisingly, 91% of employees never hit the established limits — so the company opted to reduce the cost of basic settings rather than lowering spending caps.
Routing, Cache, and Context Economy
Coinbase's internal system pre-analyzes each request and automatically routes it to the most suitable model, considering the likelihood of a cache hit and the final cost. For example, strategic planning requires an advanced model, but for routine execution, it is excessive and expensive. Armstrong insists that, ideally, the choice of model should be automated by the AI itself, not a human.
Special attention is given to caching. Cache misses are the fastest way to inflate the bill. All requests at Coinbase are configured to reuse already processed information. In the LibreChat service, the cache hit rate increased from 5% to 60% after proper configuration. "Context economy" is also critically important: Armstrong advises starting new sessions when switching tasks, strictly limiting file context, and disabling unused tools. The goal is not to spend fewer tokens in principle, but to avoid wasting them. It is this approach that allowed Coinbase to nearly halve its AI costs while consumption continues to grow.
Deutscher's "Barbell" Strategy
Analyst Miles Deutscher promotes a similar method, calling it "token engineering." He proposes a "barbell" strategy to reduce AI costs by 50% or more. Deutscher recommends entrusting the first 10% of work and project planning to the smartest models (Opus, GPT) — this is a critically important stage. The main 80% of routine tasks should be performed by a cheaper open-source model. He then assigns the final 10% and result verification to high-level models again. Deutscher has been applying this scheme for several months and considers it the best way to curb excessive AI spending.
Expert opinion: Coinbase's experience is a clear example that in an era of high computing costs, the winner is not the one who spends more, but the one who spends smarter. The transition from a "everyone gets the most powerful model" approach to intelligent routing and caching is not just about savings, but a new standard of engineering culture. Deutscher's "barbell" strategy logically complements this approach, offering a pragmatic template for any development team. The AI market is moving toward total optimization, and Coinbase is one of the prominent pioneers in this regard.