Coinbase CEO Brian Armstrong shared impressive results of AI cost optimization with the community. The company managed to nearly halve its AI expenses, even as token consumption continued to grow exponentially. According to Armstrong, the secret lies not in strict limits and bans, but in intelligent infrastructure tuning.

The key element of the strategy was multi-level request routing. Coinbase implemented a system where engineer requests are automatically directed to the most suitable model. For complex planning and architecture tasks, an advanced model is used, while for routine coding, a lighter and cheaper one is employed. Armstrong emphasizes that model selection should be automated by the AI itself, not by a human. Experiments with open-weight models such as GLM 5.2 and Kimi 2.7 through an internal gateway also yielded significant savings.

Caching and Context: Where the Extra 50% of Costs Are Hidden

Armstrong highlighted two critically important aspects: caching and context management. Cache misses are the fastest way to drive up costs. At Coinbase, the system was reconfigured to reuse already processed information. In the LibreChat service, the cache hit rate rose from 5% to 60% after proper configuration. This means that 60% of requests do not require repeated computations, directly reducing costs.

Equally important is "context economy." Armstrong recommends starting new sessions when switching tasks, strictly limiting context to files, and disabling unused tools. The goal is not to spend fewer tokens, but to avoid wasting them on excessive context. As a result, despite growing consumption, Coinbase managed to plateau its expenses and then nearly halve them.

Miles Deutscher's "Barbell" Strategy

Analyst Miles Deutscher independently described a similar strategy, calling it the "era of token engineering." His approach is a "barbell" strategy: entrust 10% of the most complex and critical tasks (planning, architecture) to the smartest and most expensive models (Opus, GPT), handle 80% of routine work with cheap open-source models, and assign the remaining 10% (final review and refinement) back to top-tier models. Deutscher claims to have been using this scheme for several months and considers it the best way to reduce excessive AI spending by 50% or more.

Analytical commentary from Cryptalist: Coinbase's experience is not just a case of savings, but a demonstration of the AI market's maturity. We are moving from the "more is better" era to the "smart optimization" era. For crypto companies, where every dollar counts, these principles are not a luxury but a necessity for survival in a highly competitive environment. The "barbell" strategy and intelligent routing will become the de facto standard for efficient AI use in 2024-2025.