The performance gap between US and Chinese AI models may be widening due to second-order effects of chip controls. By limiting inference at scale, the controls reduce the volume of customer interactions and feedback Chinese firms receive. This starves them of the data needed to identify and patch model weaknesses on diverse, real-world tasks.

Contrary to rising concerns, a joint U.S.-U.K. government report shows China's Kimi K3 model is not a frontier competitor in cybersecurity. It scored less than half of what leading U.S. models achieved on critical benchmarks, failing a network takeover attack that American models can complete.

Leading US models have safety features that block analysis of hacking tools and logs. This forces cybersecurity teams, like Hugging Face after a breach, to use less-restricted Chinese open-source models for essential forensic analysis, creating a security paradox.

Z.AI has released GLM 5.1, a massive open-source model that outperforms top US models on some coding benchmarks. Its design for 'long horizon tasks'—running autonomously for hours—signals a major advancement for China's AI ecosystem, challenging the narrative of a persistent US technological lead.

Publicly released Chinese AI models appear less capable at cyber tasks than Western counterparts. This is likely a deliberate strategy to avoid provoking their own government. However, their internal models, especially for military and state security, are undoubtedly far more advanced and closer to the cutting edge.

Despite strong benchmark scores, top Chinese AI models (from ZAI, Kimi, DeepSeek) are "nowhere close" to US models like Claude or Gemini on complex, real-world vision tasks, such as accurately reading a messy scanned document. This suggests benchmarks don't capture a significant real-world performance gap.

According to Meter, Chinese AI models are generally 9-12 months behind U.S. frontier models. Furthermore, there's a "colloquial sense" that their reported benchmark scores may overstate their true capabilities on novel, real-world problems, suggesting potential benchmark optimization.

In the vacuum left by banned US frontier models, Chinese labs are releasing powerful and cost-effective open-source alternatives like ZAI's GLM 5.2. These models are proving competitive on valuable, complex tasks like UI design and coding, but at a fraction of the cost.

Self-imposed safety pauses and regulatory hurdles on US frontier models create a vacuum. Chinese open-weight models like GLM-5.2 are now as capable as the *currently available* US versions, eroding the American lead while its most advanced models are benched, effectively ceding ground in the global AI race.