From Moonshot AI's new generation artificial intelligence model, Kimi K3, to the public beta launch of the DeepSeek-V4-Flash official API, and the official launch of Alibaba's Qwen3.8-Max on August 3rd, Chinese AI products have been rolling out in rapid succession. As the U.S. government is still trying to deal with the "black swan event" brought on by a single Chinese company, a larger-scale and longer-lasting wave of technological advancement is already surging forward.

Bloomberg reported on the 4th that these three models, along with ByteDance's Seedance 2.5 released on July 31st and Zhipu AI's GLM-5.2 launched via its Z.ai platform in June, mean Chinese AI companies have released five major models in just eight weeks. These AI models are rapidly closing the gap with Silicon Valley, "not only offering low-cost solutions but also possessing high-end capabilities such as reasoning, code generation, and complex task processing, beginning to pose a real challenge to U.S. counterparts like OpenAI." The report described this new competitive landscape created by Chinese companies as a scenario where any enterprise that cannot offer leading technology or break market prices will face elimination. This is the so-called "DeepSeek elimination threshold" — AI companies must either lower prices to match DeepSeek or invest heavily in developing stronger models, or risk being phased out.

More profoundly, the progress of these AI models is happening within China's ecosystem. As Chinese tech analyst Poe Zhao, founder of the Hello China Tech account, stated, the most significant change since DeepSeek's first debut is that China's AI development is no longer a single company's breakthrough. "The recent series of releases shows that China has formed a system capable of continuously producing models close to the global frontier." CNBC stated on August 2nd that the U.S. has, for years, underestimated China's commitment to long-term technological progress and its ability to translate domestic industrial strategy into a global competitive advantage. While the U.S. government tends to assess China's progress company by company and product by product, often dismissing each advancement as an exception or unsustainable, China has consistently pursued a patient and meticulous strategy, aiming to cultivate an environment where the entire ecosystem can innovate and deploy simultaneously.

On the 3rd, Alibaba's Qwen team officially released Qwen3.8-Max. Based on the Qwen3.5 architecture, this model scales its parameters to 2.4 trillion, with significantly enhanced capabilities in programming, office tasks, scientific research, and long-cycle tasks. According to complete benchmark test data released by the Qwen team, Qwen3.8-Match or exceeds Anthropic's Claude Fable 5 on several core indicators. This reminded U.S. media of the Kimi K3 model released by Moonshot AI two weeks prior. That model is considered to be close in performance to the most expensive AI products on the U.S. market but at a much lower budget, once again raising questions about the effectiveness of U.S. policies restricting advanced chip exports to China.

DeepSeek, the Chinese company that previously challenged U.S. AI dominance, also launched the official API for DeepSeek-V4-Flash on July 31st. On the same day, ByteDance's new-generation video creation model, Seedance 2.5, quickly swept all competitors in the AI video generation field upon release. Bloomberg noted that Zhipu AI announced its new-generation open-source large model GLM-5.2 was fully opened to users in June, once becoming the highest-ranked open-source model globally. Including this, Chinese AI companies have launched five major models in the past eight weeks. The report pointed out that this "model blitz" shows Chinese developers are approaching, and in some areas surpassing, U.S. companies long considered leaders at the global AI frontier. Moreover, the core competitiveness of Chinese companies is no longer just low prices but also extreme efficiency. Tests by the internationally renowned model benchmark platform Artificial Analysis showed that completing a complex real-world task using DeepSeek-V4-Flash costs only 3 cents, while the cost for U.S. companies is measured in dollars — $3.15 for Claude Fable 5 and $1.86 for OpenAI's GPT-5.6 Sol. Reuters interpreted that this comparison method measures value more accurately than simple pricing because it accounts for the amount of data the model must process and generate to complete a task. Even if a model has a low list price, the final cost can still be high if it requires many steps to reach an answer.

U.S. media believes this cost gap is causing substantive changes in the competitive landscape for AI clients between China and the U.S., especially in markets outside the U.S., further intensifying tech competition between the two countries. After Kimi's breakthrough, senior Trump administration officials, including Treasury Secretary Scott Bessent, threatened to impose sanctions on Chinese AI under the pretext of so-called "intellectual property theft." President Donald Trump also told the media last week that his team is weighing the necessity of restricting China's AI development against ensuring U.S. companies maintain a competitive edge. "We have to be careful on both sides. We don't want to restrict them (but also don't want to) suddenly find ourselves behind China."

In Bloomberg's view, in the competition for AI developers, U.S. companies now face two paths: either beat DeepSeek on price or surpass it in capability. A widely circulated benchmark test chart from Artificial Analysis recently showed the formation of a "DeepSeek death zone" — where, compared to DeepSeek, a company offering the same capability at a higher price, or the same price for weaker performance, will see its business prospects severely challenged. The report stated that breaking through this "danger zone" has become key to the long-term survival of AI companies. Mid-market competitors are now under increasing pressure to either lower prices to match DeepSeek or invest heavily in developing stronger models. Meanwhile, models with larger parameter scales like GLM-5.2, Kimi K3, and Qwen3.8-Max are entering a higher-performance competition tier.

Dermot McGrath, founder of ZenGen Labs, a research and strategic advisory firm focused on AI-native frontier technology, said, "I think what DeepSeek-V4-Flash really changes is the economic model for intelligent agents." McGrath explained that when researching hundreds of companies, he still uses Anthropic's Claude Code for architecture and action plans, then delegates execution tasks to DeepSeek within the same environment, where it deploys specialized AI agents to complete queries and tasks. This is because, as AI service costs have risen in recent months, improving efficiency has become a more important goal. For AI service providers, this means they don't need to completely replace the most advanced U.S. products; they just need to maximize service efficiency to exist as a complement. McGrath noted that just a few months ago, companies wouldn't have adopted such a workflow because Chinese models weren't reliable enough in tool calling and long-running intelligent agent workflows. However, Qwen's recent upgrade precisely addresses this pain point, with the new version enhancing long-cycle autonomous execution capabilities. Zhipu AI's GLM-5.2 also focuses on upgrades in code generation and long-range task execution.

It was reported that Moonshot AI and Alibaba chose a riskier strategy, building large models with over 2 trillion parameters to demonstrate sophisticated technology. Both companies also use a Mixture of Experts (MoE) architecture, activating only a portion of the model's parameters for a given task to reduce energy and computing costs. In contrast, U.S. AI giants OpenAI and Anthropic have not publicly disclosed their specific parameter scales. Parameters are seen as the neural connections of an AI's "brain," helping the system store information, process data, and generate responses. Currently, both OpenAI and Anthropic are preparing for initial public offerings (IPOs) with target valuations of at least $1 trillion. These staggering valuations are highly dependent on high-profit business models. However, the low-cost, open-source models provided by Chinese companies are challenging the pricing power of U.S. AI companies. Kai-Fu Lee, founder and CEO of 01.AI, stated that without these Chinese open-source models, OpenAI and Anthropic might easily rake in huge profits, but now the market has alternative options that are cheaper. OpenAI previously abandoned its "Sora" app and other video generation businesses due to high costs, a market gap quickly filled by Chinese companies, which have now taken a clear competitive advantage in the AI video generation field. Kuaishou's Kling AI video tool recently secured approximately $2.8 billion in funding with support from Alibaba and Tencent. Bloomberg analysts noted that a major advantage of China's AI industry is that companies and investors are willing to accept short-term return pressures to funnel resources into long-term technological competition. A company exemplifying this strategy is ByteDance, whose video AI model Seedance, besides Kling AI, is another representative model. Leveraging ByteDance's massive, sustained capital investment to keep prices low, it has become an industry leader.