Coinbase Switches to Chinese AI Models GLM and Kimi, Cuts AI Spending by 50% (mlq.ai)

Coinbase has announced a significant shift in its artificial intelligence strategy by defaulting to Chinese AI models GLM 5.2 from Zhipu and Kimi 2.7 from Moonshot AI, resulting in an impressive 50% reduction in AI spending. This change is primarily driven by a stark cost advantage; GLM 5.2 operates at $1.40 per million input tokens, making it roughly five times cheaper than Anthropic's Opus 4.8, while also outscoring it on the SWE-bench Pro coding benchmark. In addition to the model switch, Coinbase has optimized its infrastructure with an automated routing system that matches tasks with the most suitable model, achieving a 12x improvement in cache hit rate. This transition highlights a growing trend in which U.S. companies, including Snowflake and AI startup Lindy, are turning to open-weight models from China to combat rising AI costs. As enterprises face mounting financial pressure from inflated pricing by Western AI providers, Coinbase's strategy could set a precedent for similar moves within the industry, potentially undermining the pricing power of major players like OpenAI and Anthropic. However, it's important to consider the regulatory and geopolitical implications of integrating Chinese AI technologies within sensitive sectors like finance, which remain largely unaddressed in this context.