Chinese AI models now handle roughly 40% of all tokens processed on OpenRouter, up from under 2% in late 2024, according to TechTimes. The dominant model driving that shift is GLM-5.2, a 744-billion-parameter mixture-of-experts model developed by Z.ai (formerly Zhipu AI), which has been on the US Department of Commerce Entity List since January 2025 for its role in advancing Chinese military modernization through AI.

The model’s MIT license does not protect teams that access it via API from China’s National Intelligence Law. Most developer teams cannot self-host the model’s 1.5-terabyte weights.

From Rounding Error to Dominant Traffic Tier

OpenRouter routes live developer API traffic across more than 400 models from over 60 providers for more than 8 million users, roughly 47% of them American. It is not a benchmark leaderboard. Its rankings reflect what developers actually deploy in production.

Chinese AI companies held five models in OpenRouter’s daily top-50 at the start of 2025. By May 2026, that count had grown to 20, according to Our World in Data analysis of OpenRouter’s production data. By April 2026, Chinese-developed models collectively processed more than 45% of all tokens on the platform, according to Digital Applied’s analysis of that month’s rankings.

The economics of agentic coding drove the acceleration. Programming workloads grew from roughly 11% of total OpenRouter token volume at the start of 2025 to more than 50% by mid-2026. Agentic coding runs, overnight jobs that invoke a model thousands of times, account for the majority of output tokens. The unit cost per million tokens became the dominant operating expense.

Performance Close, Price Not Even Comparable

GLM-5.2 was trained entirely on roughly 100,000 Huawei Ascend 910B chips using the MindSpore framework with no Nvidia hardware at any stage, according to TechTimes. Z.ai released the model weights on HuggingFace and ModelScope under an MIT license on June 16.

On FrontierSWE, which tests open-ended technical projects measured in hours and is evaluated by independent firm Proximal, GLM-5.2 scored 74.4, within one percentage point of Claude Opus 4.8’s 75.1 and ahead of GPT-5.5’s 72.6. On agentic tool use measured by MCP-Atlas, GLM-5.2 reached 77.0, nearly matching Claude Opus 4.8’s 77.8. On SWE-bench Pro, it scored 62.1 against GPT-5.5’s 58.6, trailing Claude Opus 4.8’s 69.2 by seven points.

The gaps that persist are concentrated in the hardest evaluations. On SWE-Marathon, the most demanding long-horizon agentic test, GLM-5.2 scored 13.0 against Claude Opus 4.8’s 26.0. On ARC-AGI-2, designed to test genuine fluid reasoning, the best Chinese model scored 11.8%.

Z.ai’s API pricing runs $1.40 per million input tokens and $4.40 per million output tokens. Claude Opus 4.8 runs $5 and $25, respectively, making GLM-5.2 72% cheaper on input and 82% cheaper on output. A team running overnight agentic coding loops consuming one million output tokens daily would save roughly $7,500 per month switching from Claude Opus 4.8 to GLM-5.2’s API rate.

ZCode and the Free Agent IDE

Ten days ago, Z.ai launched ZCode, a free desktop application described as an “Agentic Development Environment.” Unlike Cursor or Claude Code, which start with a code editor and add an AI sidebar, ZCode puts the agent loop at the center. A developer types a goal, and ZCode plans the work, writes code, runs tests, and iterates. Sustained use requires a GLM Coding Plan subscription starting at $18/month, with a promotional 30% discount through September. The entry tier undercuts Cursor’s $20/month starting price.

The Sovereignty Illusion

The core tension for enterprise teams: open weights under MIT license sounds like full control. It is not. Z.ai was explicit in its architecture paper that the model’s 1.5-terabyte weight file requires infrastructure most teams do not have. Teams routing API calls through Z.ai’s infrastructure, or through third-party providers hosting on Chinese-manufactured hardware, remain subject to China’s National Intelligence Law, which compels Chinese organizations to cooperate with state intelligence work.

The Stanford University 2026 AI Index, published in April, found the overall performance gap between the best American and Chinese models had collapsed to 2.7 percentage points, but noted that figure obscures where the gap persists: in the hardest reasoning evaluations built to prevent gaming.

For agent infrastructure teams, the calculus is specific: GLM-5.2 is a viable option for cost-sensitive batch workloads where the code being processed is not proprietary. For anything else, the 82% savings comes with a data sovereignty risk that no MIT license resolves.