China has rolled out five models in eight weeks, carving out a "dead zone" that crushes mid-range competitors globally, forcing U.S. counterparts to choose between lowering prices or upgrading.
(Prior context: Why is Silicon Valley collectively anxious about China AI’s extreme value-for-price ratio?)
(Context: Nearly 200 Silicon Valley companies petition the White House: Blocking Chinese open-source models will hurt America itself)
Three cents—what can you get for that? The answer is a full set of complex real-world tasks, handled by China’s DeepSeek new model, V4 Flash. According to Bloomberg’s latest report, performing the same tasks with Anthropic’s flagship model, Claude Fable 5, would cost $3.15—more than 100 times as much.
They observed that over the past eight weeks, Chinese AI manufacturers have been pushing the ceiling of global benchmark tests one after another like a relay race, creating a "dead zone" on the market that no one dares to enter.
Five models, one production line
When you lay out the timeline:
In June, Beijing-based Z.ai launched GLM-5.2, which at the time became the highest-ranked open-source model globally.
Two weeks ago, Moonshot AI's Kimi K3 demonstrated performance comparable to America's most expensive options, using a relatively modest budget, prompting renewed questions about who exactly the U.S. chip export controls are really stopping.
Immediately after, ByteDance's video generation model Seedance launched version 2.5, pushing all competitors aside.
Then DeepSeek returns with V4 Flash, once again slashing prices to the floor.
This week, Alibaba also launched its most advanced Qwen3.8-Max, which appears to match the performance of leading flagship models.
Beijing-based technology analyst Poe Zhao told Bloomberg: “The first DeepSeek moment seemed like an exception, but this recent wave of releases shows that China now has a system capable of repeatedly producing models that are close to global cutting-edge standards.”
$0.03 vs. $3.15
Independent evaluation agency Artificial Analysis’s cost benchmark test shows that performing the same complex real-world task, DeepSeek V4 Flash costs just $0.03, while Claude Fable 5 costs $3.15. As Chinese manufacturers price their services in fractions of a dollar, their American counterparts still operate at dollar-level pricing—putting the two countries’ competition for global customers on entirely different scales.
In the death zone chart generated by Artificial Analysis, the x-axis represents cost and the y-axis represents capability. Models that are more expensive than DeepSeek without being smarter, or equally cheap but less capable, are almost certainly doomed to fail.
The three large models—GLM-5.2, Kimi K3, and Qwen3.8-Max—have moved into the high-performance zone outside the death zone thanks to their enhanced capabilities. Hugging Face engineer Tiezhen Wang also wrote on X that these three recent models have once again pushed the frontier of open-source models, with the gap narrowing significantly within just one or two months.
The recent three model releases have set a new frontier for open-source models!
The gap is closing very quickly and it only happened within a month or two.
How would this figure look by the end of this year?
Impressive and keep up with your great work! @Kimi_Moonshot… pic.twitter.com/uCVmqba4Wn
— Tiezhen WANG (@Xianbao_QIAN) July 31, 2026
The low cost, who's swimming naked?
Of course, cheap always comes at a cost. Bloomberg senior analyst Rob Lea observes that Chinese AI manufacturers are trapped in a brutal price war, prioritizing market share over profitability.
The video generation field best illustrates this: ByteDance’s Seedance offered discounts that Rob Lea described as “a 99% discount relative to market rates,” drastically lowering the market’s price floor. Meanwhile, Kuaishou’s Kling AI, backed by $2.8 billion in funding from Alibaba and Tencent, seized the market opportunity during OpenAI’s hiatus on Sora due to excessive costs.
Technical experts have also cast doubt. Micah Hill-Smith, CEO of the benchmarking platform Artificial Analysis, told Bloomberg that even as China continues to release breakthrough models, analysis still shows Chinese models are generally 3 to 9 months behind their U.S. counterparts.
On the other hand, geopolitical calculations are also unfolding in parallel. The U.S. administration has threatened sanctions citing intellectual property concerns over Moonshot AI’s technological breakthroughs; last Wednesday, Trump told reporters in the Oval Office: “We need to be careful on both fronts—we don’t want to be so restrictive that we fall behind China.”
Both OpenAI and Anthropic are currently preparing for IPOs, with valuation targets set at at least $1 trillion—but this depends on sustaining a high-margin business model. Li Kaifu told Bloomberg directly: “Without these Chinese open-source models, OpenAI and Anthropic would be laughing all the way to the bank. Now there’s another option—and it’s cheaper.”
