According to a report from Zhitong Finance APP, UBS Securities China internet industry analyst Xiong Wei believes that the cost-effectiveness competitiveness of Chinese AI models is increasingly determined by the type of task itself, with complex and long-duration tasks being particularly critical. Although overseas frontier models carry a higher per-token price, their stronger token efficiency and task success rates partially offset that price disadvantage. More developers are rolling out smaller or Flash models to optimize inference efficiency for scale-driven tasks.

UBS notes that discussions over whether to restrict open-weight models are introducing uncertainties for the overseas expansion of related models. Leading vendors are tightening anti-distillation measures, and growing attention from the U.S. policy side on industrial-scale distillation is adding to doubts about the pace of capability catch-up. The bank judges that as the most advanced models and less advanced models each take on different tiers of tasks, their pricing trajectories will diverge at an accelerating pace: competition in the scale-driven market will continue to push down per-token prices, while pricing power at the capability frontier may prove more resilient.

UBS expects Chinese AI labs will continue to scale up computing power to support training of next-generation, larger-scale models. The bank points out that future improvements in model capabilities will be driven more by diversified data sources, synthetic data, and continuously accumulated domain-specific environmental data. In the scale-driven market, Chinese and foreign developers are accelerating the launch of small-parameter or Flash models whose performance is close to flagship models but whose prices are significantly lower; global peers are also improving cost efficiency through direct price cuts, resetting free token quotas, and reducing KV cache costs.

At the capability frontier, stronger model capabilities and higher token efficiency can still lower the cost per successful task even when per-token pricing is higher, so frontier models are expected to defend or even expand their price premium. UBS believes the pace of subsequent model releases may become more cautious, but the internal R&D intensity of AI labs will not decline, and directions such as AI-assisted R&D and recursive self-improvement will remain exploration priorities. Chinese labs are expected to continue expanding computing power and to use diversified data, synthetic data, and domain-specific environmental data to drive capability upgrades. The bank also notes that progress by Chinese labs in overseas distribution, architectural innovation, and open-source R&D collaboration may not yet be fully priced in by the market.

For more heavy-weight Hong Kong stock news, download the Zhitong Finance app. For more Hong Kong stock and overseas wealth management information, please visit www.zhitongcaijing.com (search for "Zhitong Finance"); to join the Zhitong Hong Kong stock investment group, please add the Zhitong customer service WeChat (ztcjkf).