EVERYTHING
According to The-decoder, Coinbase has begun integrating Chinese artificial intelligence models into its infrastructure to optimize spending. While the company continues to process more tokens than ever before, CEO Brian Armstrong noted that these strategic shifts have allowed the firm to pay half what it previously spent on AI services.
Strategic adoption of low-cost models
The transition involves moving away from exclusively high-cost Western providers toward models like GLM 5.2 and Kimi 2.7. Although developers retain the freedom to select their preferred models, data suggests that 91 percent of users never reached their previous usage limits. This trend is not isolated to Coinbase; other major players are making similar adjustments to manage overhead.
Recent industry movements include:
- The CEO of startup Lindy recently transitioned to Deepseek v4.
- Snowflake is currently testing Chinese models as more economical alternatives to offerings from OpenAI and Anthropic.
- Coinbase has implemented an automatic routing system that selects the optimal model based on specific tasks, pricing, and caching potential.
One of the most significant technical improvements reported by Coinbase involves their caching strategy. By improving caching alone, the company successfully pushed its hit rate from 5 percent to 60 percent. To support this efficiency, developers are instructed to maintain lean contexts and initiate fresh sessions for new tasks, a practice categorized as context engineering.
Pricing pressure on Western AI labs
The shift toward cheaper alternatives creates a difficult environment for Western AI laboratories, particularly those seeking initial public offerings. These companies must now prove they can hit growth targets while facing a price war from both domestic and international competitors. Reports suggest that OpenAI and Anthropic are already entering a pricing conflict, with OpenAI offering various variants of its 5.6 models at different price points to compete with Claude's efficiency.
Coinbase also maintains a policy of transparency regarding developer usage without imposing hard caps. This mirrors the "tokenmaxxing" trend seen at companies like Amazon and Meta, where high token consumption was previously celebrated. However, Armstrong has introduced a layer of accountability to this practice. "The more you spend on AI, the more impact we expect," — Brian Armstrong, CEO of Coinbase. By combining aggressive cost-cutting with performance monitoring, the company aims to balance rapid technological adoption with fiscal responsibility.