Coinbase CEO Brian Armstrong disclosed in early July that the US cryptocurrency exchange has switched its default AI models to GLM 5.2 from Zhipu, a Chinese lab, and Kimi 2.7 from Beijing-based Moonshot AI. The shift cut Coinbase’s internal AI spending nearly in half despite increasing token consumption to record levels.

GLM 5.2 costs approximately $1.40 per million input tokens and $4.40 per million output tokens. Anthropic’s Opus 4.8, the model Coinbase previously relied on, costs $5 and $25 respectively. The Chinese model is roughly five times cheaper for comparable performance.

Performance vs Cost

On the SWE-bench Pro coding benchmark, GLM 5.2 scored 62.1, outperforming OpenAI’s GPT-5.5 at 58.6. Coinbase is not sacrificing capability. It is accessing better performance at lower cost. The math is simple: why pay more when the alternative is superior?

This performance gap signals that Chinese AI development has matured. Zhipu is competitive with Anthropic and OpenAI. The gap closed faster than expected in 2026.

The Self-Hosting Strategy

Coinbase downloaded the open-weight model to its own servers for self-hosted operation. Code and queries do not flow to servers in China. Coinbase retains data privacy while accessing the model’s capabilities. This structure sidesteps some regulatory concerns.

But Armstrong did not address potential regulatory risks of routing US financial-sector workloads through Chinese-developed AI models. Regulators and competitors may press on this point.

What This Signals

US tech giants are quietly switching to Chinese AI. Lindy moved 100% of its API traffic from Anthropic to DeepSeek V4. These are not experiments. These are production decisions by real companies optimizing margins.

The trend signals three things: Chinese AI is competitive, cost matters more than brand loyalty, and open-source models level the playing field. Proprietary APIs from Anthropic and OpenAI face pricing pressure they haven’t faced before.

Coinbase’s move is a business decision, not a political statement. But it signals a shift in US AI economics that regulators cannot ignore.