Coinbase now runs most of its AI workloads on Chinese open-weight models, according to CEO Brian Armstrong. The company deployed GLM 5.2 and Kimi 2.7 alongside an autonomous routing system that selects models based on task complexity, cost, and caching potential. The result: a 50% reduction in AI spending despite rising token consumption, Armstrong said on X.
How the Routing System Works
The routing layer matches each request to the cheapest model capable of handling it. Developers can still manually select any model, but 91% never hit their previous usage limits anyway, according to The Decoder. The system runs autonomously, requiring no per-request human intervention.
Caching improvements proved equally significant. Coinbase pushed its cache hit rate from 5% to 60% by enforcing lean context windows and fresh sessions for new tasks. That alone eliminated a large share of redundant model calls.
A Pattern, Not an Outlier
Coinbase is not the first enterprise to make this move. Lindy CEO Flo Crivello switched entirely to DeepSeek, saving millions in the process. Snowflake CEO Sridhar Ramaswamy has publicly stated that GLM 5.2 competes with Opus 4.7 at a fraction of the cost, according to The Decoder.
The pattern is consistent: enterprises are discovering that most of their AI workloads do not require frontier models. When a cheaper model handles the task adequately, routing to it automatically is pure margin.
Accountability as a Differentiator
Armstrong added one constraint that separates Coinbase from the “tokenmaxxing” trend reported at Amazon and Meta. “The more you spend on AI, the more impact we expect,” Armstrong said. Usage is visible per developer but not capped, creating accountability without restricting experimentation.
The Pricing Pressure on Western Labs
The enterprise shift to Chinese models adds direct pricing pressure on OpenAI and Anthropic at a sensitive moment. Both companies are approaching potential IPOs, and their revenue growth narratives depend on enterprise API consumption, The Decoder reports. A price war between the two labs may already be forming, with OpenAI’s GPT-5.6-Sol priced at GPT-5.5 levels while claiming better token efficiency than Claude Fable and Mythos.
For teams building agent infrastructure, Coinbase’s playbook is worth studying: autonomous routing, aggressive caching, and accountability metrics. The cost reduction didn’t come from switching to inferior models. It came from matching model capability to task requirements automatically, then measuring whether the spend produced results.