Will China Win the Proxy AI War Against the US? Australian Think Tank: Can 'Crush' Western Models at Ultra-Low Cost

Will China Win the Proxy AI War Against the US? Australian Think Tank: Can 'Crush' Western Models at Ultra-Low Cost

### The Growing Gap in AI Performance and Pricing Between China and the US

As global artificial intelligence (AI) technology advances, the gap between China and the United States in AI capabilities and computing power is drawing increasing attention. While U.S. Treasury Secretary Scott Bessent has accused China of stealing AI models and claimed the U.S. leads China by at least a year, the Australian Strategic Policy Institute (ASPI) issued a new report on July 21 warning that China's low-cost AI models could pose a major challenge to Western models in the upcoming era of Agentic AI.

The report, authored by analyst Alex Colville, highlights that within just one month, Chinese AI startups Z.ai and Moonshot have launched models—GLM-5.2 and Kimi-K3, respectively—that match the intelligence level of U.S. rivals like OpenAI and Anthropic, but at a fraction of the cost. GLM-5.2 is already being adopted by Silicon Valley startups, and Kimi-K3 is poised to gain significant market traction.

### The Cost Crisis in Agentic AI

Agentic AI, which performs complex, multi-step tasks autonomously, consumes thousands of times more tokens than standard language models. According to recent estimates from MIT, Stanford, and Google DeepMind, Agentic AI tasks require about 3,500 times more tokens than simple inference tasks. A single coding or research task can consume millions of tokens, with heavy users spending tens or even hundreds of millions per day.

China's models offer a stark cost advantage. Z.ai charges just $1.92 per million output tokens for GLM-5.2, compared to Anthropic's $25 for Opus 4.8—a difference of over 12 times.

### Global Shift Toward Chinese AI Models

This pricing disparity is already influencing real-world adoption. Microsoft’s decision to discontinue its Claude Code subscription was partly due to high token costs, and the company is now reportedly considering integrating China’s DeepSeek to power its Copilot Cowork assistant.

British tech entrepreneur and writer Azeem Azhar has integrated Chinese Agentic AI into his research operations, adopting Xiaomi’s MiMo-2.5-Pro due to its significantly lower token costs—critical for his team, which uses over 100 million tokens daily. As Azhar noted, 'At the scale of Agentic AI, even tiny cost differences accumulate into massive budget gaps.'

### Can Western AI Pricing Be Sustained?

The report warns that current pricing models at OpenAI and Anthropic may be 'unsustainable' in the Agentic AI era. In contrast, DeepSeek has permanently slashed its API token prices by 75%, making its services potentially 1/34th the cost of OpenAI or Anthropic.

Independent AI evaluator Artificial Analysis tested 657 Agentic AI tasks across models. Running these on Anthropic’s Opus 4.8 cost nearly $1,000, while Z.ai cost $270. DeepSeek-V4 Flash, despite consuming the most tokens, completed all tasks for just $14. Kimi-K3 outperformed all others in task completion accuracy and cost only about one-third of Anthropic’s Fable 5.

### The Rise of Chinese AI and Future Challenges

While Chinese models still face limitations, the report suggests 'this may be changing.' Once Chinese AI achieves both high performance and low cost, Western models will face severe competition.

Moreover, if high costs are a burden even for well-funded Western firms, they will be insurmountable for many 'Global South' nations. The report urges Western companies to balance technological advancement with affordability to remain competitive.