Three stories ran through the Chinese AI press in the same week, each one circulating on its own channel. The first was a coordinated price cut by the three American frontier labs. Within nine days, OpenAI, Google and Anthropic each slashed headline pricing on their main models and arrived at the same pair of numbers: USD 2 per million input tokens and USD 10 per million output tokens.
The second was a claim that 55 per cent of China’s AI token consumption goes into AI short dramas. The third was a vision model identifying a company founder as a well-known internet comedian, confidently and repeatedly.
Each claim stands on its own. Put together, they contradict each other. If Chinese models really only made short dramas, American labs would not need to cut prices in the same week. If the price advantage were strong enough to force rivals into loss-making bids, a 55 per cent consumption mix would not be evidence of backwardness.
Nine days, one price point
On 22 September, OpenAI released GPT-6 Sol and GPT-6 Luna. Sol was priced at USD 2 input and USD 10 output per million tokens. Luna came in at USD 0.10 and USD 0.50. The official comparison was GPT-5.6 Sol at a promotional USD 4 and USD 20, so the new flagship halved the old one.
Google followed on 1 October. Gemini 4 Argon quoted the same USD 2 and USD 10, and pushed the single-response ceiling from 64,000 tokens to one million. Anthropic’s Claude Sonnet 5.5 then landed on the identical pair.
On 7 October, Anthropic released Haiku 5.5 at USD 0.10 input and USD 0.50 output for requests under 100,000 tokens. That is one tenth of Haiku 4.5 and exactly level with OpenAI’s GPT-6 Luna. The same day, OpenAI put GPT-6 into ChatGPT, with Sol behind the paid tier and Luna on the free and Go tiers.
None of the announcements explained the timing. The reason is not hard to find. Two years of Chinese open-weight models have pushed the global price band down, and the third-party evaluator Artificial Analysis quantifies the gap: a single benchmark run on DeepSeek V4 Flash costs about USD 0.03, against USD 3.15 on Anthropic’s flagship Claude Fable 5.
That is a 100-fold difference, and developers responded. A term has even appeared in American developer circles, Chinamaxxing, meaning the practice of swapping every swappable pipeline over to a Chinese model. One founder calculated that scaling from 5,000 to 50,000 paying users would burn the whole round in three months, so the model swap came before the growth plan.
The discount is mostly temporary
The headline cuts are largely time-limited. Google’s promotional pricing runs until 30 November and then doubles to USD 4 input and USD 20 output. Gemini 3.8 Flash doubles on 1 January 2027. Haiku 5.5 carries a 100,000 token line, above which the rate jumps five-fold.
OpenAI’s product lead has said GPT-6 is a permanent price rather than a promotion. The internal trend still points up at the top of the range. Flagship Astra lists at USD 10 input and USD 50 output. In July 2025 the GPT-5 output price was USD 20, and in March 2026 GPT-5.4 was USD 15.
So the flagship tier has been getting more expensive while the mid and low tiers came down. What was cut is the sticker price, not the underlying cost of compute. The tiers that could be cut are the ones whose margin was thin to begin with.
Where the 55 per cent comes from
The claim that 55 per cent of Chinese token consumption goes to short dramas traces back to an institutional investor briefing by Chinese Online on 3 June. It is a single company’s disclosure, not a statistical agency figure and not an independent survey. The company did not publish its method, which platforms it covered or how it converts tokens across modalities.
The most widely shared version also changed the unit. The original statement combined AI short drama and video generation at 55 per cent. It was compressed in circulation to short drama alone at 55 per cent.
The physics of video also gets lost. Under 720p, a five-second shot consumes roughly 100,000 tokens, about the same as a 100,000-word script. A 90-minute AI short drama runs to about 110 million tokens for the finished cut alone, before the industry norm of three or four retakes per shot, which pushes a single title past 100 million tokens comfortably. Text sits in the hundreds of thousands, video in the hundreds of millions. The gap is 1,000 to 10,000-fold, and it reflects the compute density of video rather than a preference for soap opera plots.
The number that gets left out is the total. OpenRouter weekly data shows Chinese model calls at 12.96 trillion tokens in the week of 30 March to 5 April, against 3.03 trillion in the United States, more than four times the American figure. By the week of 5 to 11 October, China was at 86.74 trillion, up 50.96 per cent week on week and above the United States for 24 consecutive weeks, with all of the top five models globally Chinese.
Put the four-fold multiple next to the 55 per cent and the conclusion inverts. Chinese token spend on short drama alone approaches twice total American consumption. Even removing short drama entirely, the remaining 45 per cent multiplied by four is still about 1.8 times American consumption. Both numbers come from the same data set. Only the ratio travelled.
The compute constraint underneath
The more useful question is what the models can actually do, and what it costs to keep them running. Tencent has acknowledged publicly that overall compute is short and is slowing Hunyuan’s iteration. DeepSeek’s V4 Flash became effectively unavailable through its API in August when traffic beat forecasts, and the company then raised API prices citing compute pressure. Kimi K3 paused new consumer subscriptions to add capacity. Zhipu raised API prices 83 per cent in the first quarter and saw call volume rise 400 per cent anyway.
Chinese labs also still run two separate pools. Training, the heavier workload, depends on whatever Nvidia hardware they can obtain. Inference is gradually moving to domestic cards. Huawei rotating chairman Xu Zhijun put the constraint plainly: what limits Ascend volume today is no longer demand, it is supply.
That is changing. Huawei’s Ascend 384 supernode had more than 750 units in commercial service at WAIC in July, and by its September partner conference the 910C supernode had passed 1,000 deployments, with 650 million cumulative online card-hours and more than 40 models natively pretrained on Ascend. On 9 July, the Greater Bay Area’s first Ascend 10,000-card cluster built for training domestic models on domestic silicon went live in Shaoguan with 11,520 cards and 9,000 petaflops, from groundbreaking to production in 90 days. Meituan said its trillion-parameter LongCat-2.0 was trained end to end on a 50,000-card domestic cluster. Zhipu has built a data centre planned at about 1 gigawatt on domestic chips, partly operational. The Ascend 950 supernode, shown as real hardware at WAIC, runs 1,024 cards at 1 exaflop.
The vision failure that started this week’s chatter is the least important part of the story. A model that reads images rather than seeing them will keep misidentifying people whose public images are scarce, and that will not be fixed by a price cut. What will be fixed by price is who can afford to run agents at scale, and on whose silicon.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1093260209_122189.