Coinbase Defaults Engineers to China's GLM-5.2 as US Export Controls Squeeze Anthropic and OpenAI

A Frontier Model Built on Domestic Chips

Z.ai, the Beijing-based lab formerly known as Zhipu AI, released GLM-5.2 on 13 June 2026, just one day after the Trump administration ordered Anthropic to disable its Fable 5 and Mythos 5 models globally under a sweeping export control directive. The timing turned the 744-billion-parameter open-weight model into the most discussed Chinese AI release since DeepSeek's debut. GLM-5.2 uses a Mixture-of-Experts architecture with roughly 750 billion total parameters but only about 40 billion active per token, keeping inference cheap. Z.ai claims it was trained entirely on Huawei Ascend accelerators with no Nvidia hardware, ships with a one-million-token context window and up to 131,072 output tokens, and is released under a permissive MIT licence with weights freely downloadable from Hugging Face.

Benchmark Performance and Pricing

On FrontierSWE, a long-horizon coding benchmark, GLM-5.2 trails Anthropic's Opus 4.8 by roughly one percentage point while outperforming OpenAI's GPT-5.5. Its API pricing of $1.40 per million input tokens and $4.40 per million output tokens undercuts Opus 4.8's $5 and $25 by a wide margin, making it roughly five times cheaper. The model is the highest-ranked open-weight entry on the Artificial Analysis Intelligence Index, scoring 51 on version 4.1 of the benchmark and placing fourth overall globally.

Armstrong's Cost Argument

Coinbase CEO Brian Armstrong amplified the cost case on 27 June, describing how Coinbase uses an internal LLM gateway to default engineers to open-weight models including GLM-5.2 and Moonshot AI's Kimi 2.7. According to The Information, Coinbase has cut its AI spending nearly in half even as token usage rises. Armstrong has argued that 80 percent of workloads will run on 99 percent cheaper models within 12 to 18 months. The trend is visible on OpenRouter, where Chinese models now account for over 60 percent of token traffic, up from under 2 percent a year earlier.

Security Questions Loom

The enthusiasm collides with concerns about China's National Intelligence Law, which compels Chinese organisations and citizens to support national intelligence efforts. Z.ai's cloud API falls under that law, raising data routing concerns for enterprises handling sensitive information. In May, US House lawmakers opened a formal inquiry into the cybersecurity risks of Chinese-origin AI models in critical infrastructure, naming Zhipu alongside DeepSeek, MiniMax, and ByteDance. As US export controls constrain access to American frontier models, the open-weight alternatives filling the gap arrive with their own risks that enterprises must weigh against the savings.