AI competition shifts from raw power to token cost and efficiency
Washington is considering a capability rule that would permit US models, whether open or closed, when they are no more capable than China’s best . , which uses a powerful model’s outputs to train a smaller and cheaper one, is making the policy debate more urgent. Anthropic has accused Alibaba of copying its models at an industrial scale and framed the issue as a $6 billion annual problem, though critics see that claim as an exaggerated warning.
Zhipu, which openly released GLM-5.2, argues that broad participation makes AI safer than closed barriers do. At the same time, leading AI companies are competing more directly on usage cost rather than raw ability alone. GPT-5.6 uses fewer tokens and keeps its across multiple exchanges, while Grok 4.5 is described as twice as efficient.
On 525 clinical tasks, outperformed at one-seventh of the price.
Key points
- GPT-5.6 is designed to use fewer tokens and preserve across multiple exchanges.
- Grok 4.5 is described as twice as efficient.
- beat on 525 clinical tasks at one-seventh of the price.
- The US is considering using China’s best as a capability line for permitting domestic models.
- Rules prompted by and model-copying disputes could affect the supply of affordable .