Symbolic imageUS AI figures call Chinese open-weight models safer as price gap widens
Prominent American AI researchers are publicly arguing that Chinese open-weight models are safer to use than closed US systems. The debate comes as DeepSeek's V4 Flash undercuts frontier pricing.
What happened
- At the Agentic AI Summit in Berkeley on Saturday, 1 August, AI researcher Andrew Ng called open-weight models safer than closed-weight ones.
- Ng and a colleague used Moonshot AI's Kimi K3 and Zhipu AI's GLM-5.2 to security-review OpenWorker, their open-source agent tool released last month.
- Leading OpenAI and Anthropic models had refused that security task, according to Ng.
- OpenAI's GPT-5.6 and Anthropic's Fable 5 carry added safeguards that users say also block legitimate cybersecurity requests.
- Alibaba released Qwen3.8-Max with downloadable weights on Monday, 3 August.
The view from outside
The Hong Kong based South China Morning Post frames the episode as a direct challenge to closed-source safety claims, a framing absent from the brief Chinese state coverage.
What could happen next
- Global Times' 'Ferrari at bicycle prices' billing rests on a teaser blurb; benchmarks and pricing for V4 Flash are not detailed.
- A research firm's estimate that V4 Flash costs over 100 times less than Anthropic's Fable 5 remains unverified here.
- Open: whether US developers loosen safeguards for security work, and whether cheap open weights pull enterprise deployment away from closed systems.
How the camps see it
China: Chinese labs and state media present downloadable weights and low prices as the decisive advantage, arguing scaled deployment matters more than a capability edge.
United States: US closed-source developers such as Anthropic have long held that open-source models pose a threat to society and keep safeguards that refuse risky requests.