Author: Rhythm
On October 8, Anthropic released Haiku 5.5. For short requests of no more than 100,000 tokens, it charges $0.1 per million input tokens and $0.5 per million output tokens, only one-tenth that of the previous generation, Haiku 4.5.
That morning, Dragonfly managing partner Haseeb Qureshi posted two Artificial Analysis scatter plots on X, with the caption:
"If you want the cheapest LLMs, you should now buy American."
The horizontal axis of the scatter plots is the cost of completing a benchmark task, and the vertical axis is the score. The dashed line connecting all optimal solutions is called the Pareto frontier in microeconomics. Every point on the line means that at the same cost, you cannot find a smarter model.
On the June chart, the cheapest end of the dashed line was Xiaomi MiMo, DeepSeek V4 Pro, MiniMax-M3, and Zhipu GLM-5.2, almost all names of Chinese large models.
By October 8, the entire line had been taken over by Anthropic and OpenAI. Among Chinese models, only MiMo still barely clung to the edge, while Zhipu GLM-5.3-Flash and DeepSeek V4.1 Flash were pushed to the lower right of Haiku 5.5, meaning that by comparison they cost more and scored lower.
At the Hong Kong stock market open that day, large-model concept stocks fell in response. At 10:26 a.m., MiniMax's decline widened to 10%, and Zhipu fell 5.6%. By the midday session, the Hang Seng Tech Index was down 1.93%, with the two large-model upstarts underperforming the broader market by nearly five times.
Over the past two years, Chinese open-source and high-cost-performance models had welded a kill line into the global large-model market. As long as prices were pushed into the mud, overseas closed-source models in the volume tier would struggle to sustain themselves. Star Silicon Valley companies lined up to replace their underlying pipelines with Chinese open-source large models.
Now, the table has been flipped.
The kill line has been killed.
The formation of the kill line
When GPT-4 first launched in March 2023, the official price for the 8K context version was $30 per million input tokens and $60 per million output tokens.
Intelligence was extremely scarce, supply was monopolized by a very small number of Silicon Valley oligarchs, and pricing was entirely a seller's market. That was a honeymoon period that believed "intelligence can command a premium on its own."
What shattered this myth was a technical team incubated by a Chinese quantitative hedge fund. On May 6, 2024, DeepSeek launched V2. With 236 billion total parameters and only 21 billion activated per token; relying on the MLA architecture and extreme engineering trimming, the team cut the KV cache by 93.3% and raised throughput to 5.76 times. The saved compute was directly converted into commercial pricing: 1 yuan per million input tokens and 2 yuan per million output tokens.
Domestic major tech companies were almost dragged into the price war by force.
On May 15, ByteDance's Volcano Engine released Doubao, with its main model offering an enterprise price of 0.0008 yuan per thousand tokens, claiming to be 99.3% cheaper than the industry average. On May 21, Alibaba Cloud announced cliff-edge price cuts for the Tongyi Qianwen series, with Qwen-Long charging only 0.5 yuan per million input tokens; hours later, Baidu directly announced that two main Wenxin models would be free.
By August, DeepSeek further introduced hard disk caching technology, charging only 0.1 yuan per million input tokens for cache hits.
At that time on the other side of the ocean, the giants did not fully follow suit. In July 2024, OpenAI released GPT-4o mini, lowering input and output to $0.15 and $0.6; but four months later, Anthropic launched Claude 3.5 Haiku, priced at four times the previous generation. Dario Amodei's logic at the time was that the model had become smarter, so pricing had to reflect the advancement of intelligence.
What truly made Silicon Valley feel a bone-chilling cold was January 2025.
DeepSeek-R1 burst onto the scene, with reasoning performance approaching OpenAI o1, while output cost only $2.19 per million tokens, whereas o1 was priced at $60 at the time.
On January 27, Nvidia's market value evaporated by nearly $589 billion in a single day, setting a record in U.S. stock market history. Marc Andreessen called it the Sputnik moment in AI, and Microsoft CEO Satya Nadella remarked on social platforms that the Jevons paradox had struck again.
The kill line took shape from then on.
Open weights made this blade even sharper. Facing closed-source models, customers could only do arithmetic among several monopoly price lists; once weights were open, developers could both deploy privately on their own clusters and run the same model on any third-party inference cloud. The same model had multiple suppliers, and customers also gained the option of self-deployment, shrinking the room for model providers to maintain high premiums.
A group of U.S. companies were the first to cross the river.
In October 2025, Airbnb CEO Brian Chesky publicly said that the company's intelligent customer service Agent relied heavily on Alibaba's Qianwen, and that the entire system orchestrated 13 models. Although it also integrated OpenAI's latest flagship, it was rarely called in the production pipeline because there were more cost-effective alternatives.
That same month, Cognition launched SWE-1.5, claiming it was built on "a certain industry-leading open-source base." Zhipu later confirmed that the base was GLM-4.6.
Soon after, a16z partner Martin Casado told The Economist that among U.S. AI startups pitching with open-source model architectures, about 80% used Chinese bases.
In March of this year, code editor Cursor released Composer 2, priced at $0.5 per million input tokens and $2.5 per million output tokens, and the team eventually admitted that the base was fine-tuned from Moonshot AI's Kimi K2.5.
Developers were voting with code and budgets. A report released by Mozilla showed that the token share of Chinese open-weight models on OpenRouter rose from less than 2% at the end of 2024 to more than 45% by April 2026.
In the ecological niches of volume and structured calls, the proud closed-source models once had no ability to fight back.
The cost
Killing opponents does not necessarily mean living comfortably oneself.
In the same Mozilla report, there was another set of comparative data. From May to September 2025, open models accounted for about 20% of model usage on the OpenRouter platform, but received only about 4% of model-layer revenue. This is a sample of a platform in a specific period and cannot represent the global market, but the gap between usage share and revenue share is already very obvious.
Creating value has never been equal to capturing value.
In January of this year, Zhipu and MiniMax rang the bell one after another. The interim results subsequently disclosed allowed the outside world to see the costs and losses behind revenue growth.
Zhipu's revenue in the first half of the year was 954 million yuan, of which revenue from the open platform and API business was 825 million yuan, accounting for 86.5%. The gross margin of this business had turned positive from negative, rising to 24.6%; but the company as a whole still recorded a net loss of about 2.072 billion yuan, narrowing by 12.1% year on year. Selling more tokens had begun to generate gross profit, but there was still a distance to covering all of the company's expenses.
MiniMax's revenue in the first half of the year was about $117 million, up 283.1% year on year; net loss was about $358 million, narrowing by 11% year on year. After excluding share-based payment, changes in the fair value of financial liabilities, and listing expenses, the adjusted net loss was about $293 million, expanding by 111.2% year on year.
After listing, the growth brought by low prices needs to withstand another test. How much gross profit can each call leave behind, and how much R&D and operating expense can that gross profit cover. Revenue is growing quickly, but the company as a whole is still far from profitability.
In late spring of this year, the Agent wave pushed compute consumption to a critical point.The open-source project OpenClaw swept through the global developer community, surpassing a cumulative 8.52 trillion tokens consumed on OpenRouter by early March, dominating the top spot. Within a single week in mid-March, Chinese models ran 7.36 trillion tokens, a month-over-month surge of 56.9%.
GLM-5, launched on February 12, briefly topped the call rankings, and the surging demand instantly overwhelmed infrastructure quotas. Overseas and domestic developers began complaining intensively about second-level latency and high-frequency rate limiting on the GLM Coding Plan—the first time a leading Chinese AI team publicly issued an urgent call for compute.
On February 23, Zhipu's stock price plunged nearly 23% in a single day, evaporating over HK$70 billion in market value in one day.
What followed was a round of defensive price increases.
Zhipu raised rates when GLM-5 was launched, and in March the price of GLM-5-Turbo rose another 20%, making it an average of 83% more expensive than the previous generation's specifications.
DeepSeek's V4-Pro, released on April 24, has 1.6 trillion total parameters and offered a 75% peak discount upon launch. The fire first spread to peers, with MiniMax falling about 9% and 10% over two consecutive trading days.
But by mid-August, DeepSeek also shifted to time-of-day pricing, with V4-Pro peak-hour output climbing to $3.96 per million tokens, and cache-hit costs rising more than 12-fold.
In July, Moonshot AI released Kimi K3, with API pricing pushed up to $3 per million input tokens and $15 per million output tokens—more than three times as expensive as K2.6.
Yet, unexpectedly, the price increases did not trigger severe customer churn.
Zhipu CEO Zhang Peng publicly revealed that after an 83% price adjustment in the first quarter, API call volume instead surged 400%. By the end of August, Zhipu's MaaS platform annualized revenue reached $1.6 billion, and the API business gross margin miraculously turned positive to 24.6%.
During the months-long window when compute was in short supply, Chinese model teams touched precious pricing power for the first time.
However, the kill line that once kept overseas giants awake at night was also, imperceptibly, pushed higher by themselves.
Blitzkrieg
The other side of the ocean completed a strategic pivot in the summer.
At the end of June, Anthropic released Sonnet 5, listing a launch promotional price of $2 per million input tokens and $10 per million output tokens, having originally explicitly stated that prices would return to $3 and $15 in September.
Then in July, OpenAI launched the GPT-5.6 series, and three weeks later suddenly launched a blitzkrieg, cutting Luna's price by 80%, with the input side dropping to $0.2 and the output side to $1.2.
OpenAI attributed this price cut to pure underlying engineering refactoring: rewritten GPU kernels reduced end-to-end serving costs by about 20%, and retrained speculative decoding draft models accelerated generation throughput by more than 15%.
On August 10, Anthropic withdrew its price increase notice and announced that Sonnet 5 would permanently lock in the low price.
On September 22, Anthropic brought out Opus 5.5, cutting comprehensive usage costs by 40%; on the same day, OpenAI launched GPT-6 Sol and Luna, with Luna locking its entry price at $0.1 input and $0.5 output. What followed was Haiku 5.5.
Anthropic, which two years ago firmly believed that "intelligence must command a separate premium," tore off its own label with its own hands.
The rhetoric at large model launches has changed too. No longer purely listing narrow MMLU victories, they began repeatedly emphasizing output per dollar, latency, and the engineering efficiency of each step of inference.
Moreover, alongside the debut of Haiku 5.5, Anthropic announced that it would comprehensively allocate API credits to subscribers, gifting $100 to $200 per month at the Max tier, with enterprise Team users eligible for up to $500 in offsets.
This move struck straight at the heartland. Discounts on the price list only affect single-call decisions, but once a customer's code engineering and Agent orchestration complete adaptation, the barrier to switching becomes as high as a city wall.
The American giants dare to launch this price war because of their unfathomably deep capital pools and supply chain privileges.
Anthropic's annualized operating revenue soared from about $9 billion at the end of 2025 to more than $65 billion by the end of July this year. The $65 billion financing completed in May directly pushed its valuation to $965 billion.
On the compute front, Anthropic announced in April an expansion of cooperation with Google and Broadcom, obtaining multi-gigawatt-scale next-generation TPU compute capacity, expected to come online progressively starting in 2027.
By contrast, Zhipu struggled within the year to raise more than HK$70 billion through placements and convertible bonds—less than one-seventh of Anthropic's single round of ammunition when converted to US dollars.
Those who can afford to pay are still only the super giants.
Secondary market reckoning has never been merciful.
Zhipu listed on the Hong Kong stock exchange on January 8 at an issue price of HK$116.2. The spring compute shortage failed to stop the bulls' frenzy; on June 22 its stock price surged to HK$2,980, with market value briefly touching the HK$1.33 trillion ceiling.
Then came a long decline.
In July, restricted shares were unlocked, and two rounds of placements hammered the private placement price from HK$1,588 all the way down to HK$714. On July 16, the day after Kimi K3 debuted, Zhipu fell 28.48% for the full day. On September 22, with Opus 5.5 and GPT-6 attacking in tandem, Zhipu's decline again exceeded 10%.
On September 25, Zhipu hit an intraday low of HK$610.5, with total market value shrinking to around HK$300 billion, a drawdown of nearly 80% from its historical high.
MiniMax's trajectory was equally brutal. Falling from a peak of HK$1,330, it plunged 17.98% on the unlock day and closed at HK$216 in mid-July. Nearly 60% of its first-half revenue depended on overseas customers, and overseas was precisely the battlefield first hit by the American giants' price-cut storm.
Valuation metrics also began to tighten. According to media reports, Jefferies in a September research note lowered the 2026 forecast annualized revenue valuation multiple corresponding to Zhipu's cloud business from 50x to 30x, a reduction of 40%.
Chinese models once snatched food from the jaws of giants by relying on extreme cost-effectiveness, and now this path has also been paved on the other side of the ocean. Developers come because it is cheap, and naturally leave because something else is cheaper.
The Ghost of $0.1
On August 25, 2006, on a sweltering beach in Cabo San Lucas, Mexico, Jeff Barr typed out an article introducing the Amazon EC2 beta.
At the end of the technical description, he wrote down a number destined to enter business history: developers renting a virtual compute instance would be charged $0.1 per hour.
The rest of the story is well known. AWS proactively cut prices more than a hundred times over the following decade or so, and in 2014 alone S3 storage rates were halved by 51%.
Cloud computing never achieved its empire by selling the same compute at ever higher prices. Every bit of marginal profit squeezed out by scale, in-house chips, and extreme operations was continuously returned to the end market, thereby attracting more developers. The continuously price-cutting AWS ultimately grew into a cash cow with annual revenue of nearly $130 billion and an operating margin as high as 35.4%.
Price wars are the final encirclement tactic of scale monopolists.
Those today who can push model prices down to a tenth and fill the shelves of the world's three major public clouds immediately upon release are not only selling tokens, but also absorbing tokens. While slashing API prices, Anthropic simultaneously feeds massive inference compute back into Claude Code, Cowork, and its own end-to-end Agent products.
Chinese large model teams still have not laid down their weapons. In September, DeepSeek again cut Flash series rates, Xiaomi pushed MiMo-Flash to the limit of 1 yuan per million tokens, and Zhipu also priced GLM-5.3-Flash stubbornly within a thin-margin range. But the rules for chips on the table have already been rewritten.
Low price itself is no longer an exclusive moat card.On that scatter plot, updated in real time by Artificial Analysis, not a single Chinese company's name is missing. GLM, DeepSeek, Kimi, Qwen, and MiniMax are all still there.
It's just that they can no longer draw that kill line.