Roughly 750,000 words of AI output costs $50 if you buy it from Anthropic’s Fable. The same volume from DeepSeek’s V4-Pro costs about 87 cents, Fortune reported. Z.AI’s GLM-5.2 charges $4.40. Moonshot’s Kimi K3, pricey by Chinese standards, charges $15.
Corporate America has noticed. Companies spent a year competing to burn tokens. Now they shop for the cheapest model that can finish the job, the Wall Street Journal reported. The badge of honour used to be tokenmaxxing. The new one is thrift-maxxing.
The first response to the bills was rationing, when firms capped what staff could spend. The second is substitution, and it is a far bigger problem for OpenAI and Anthropic.
“It’s like driving a Lamborghini to go to the grocery store to pick up milk when that was designed to be raced around a track,” said Mike Saeks, a field chief technology officer at Cursor who advises companies on their AI returns.
He has the receipts. Cursor priced up building a web browser from scratch. Running the whole job on OpenAI’s GPT-5.5 cost a little over $10,000. Splitting it between Cursor’s own Composer model and Anthropic’s Opus 4.8 cost $1,339.
The orchestra, not the soloist
This is not companies abandoning American labs. It is companies demoting them to the parts of the job that justify the price.
Telnyx, which builds infrastructure for AI agents, ran 1,000 of them on a top Anthropic model under a $200-per-employee subscription. Then Anthropic stopped allowing third-party operating systems on subscriptions, treating it as a breach of its terms. Paying per use instead would have cost “like 100 grand per day”, chief executive David Casem told the Journal.
So Telnyx rebuilt the stack. Z.AI models now run its 1,400 agents. Anthropic’s Fable acts as the conductor and plans the work. Open-weight models handle the implementation. OpenAI’s Sol reviews what they produce.
The legal AI startup Harvey trained GLM-5.2 itself. It then gave the model a button to call Fable 5 when a task looks genuinely hard. “We work with all of them,” said president Gabe Pereyra. Zoom’s chief technology officer Xuedong Huang reached for an old Chinese story. In it, three ordinary people combine their wits to match one genius. That, he said, is the secret sauce.
“It really feels like a bloodbath”
The labs are reacting the way any supplier does when switching costs collapse. They are paying customers to stay.
“There’s zero loyalty that I’m seeing,” said Marty Kausas, chief executive of the customer support platform Pylon. “It really feels like a bloodbath right now.” Pylon has been handed months of unlimited free usage. Kausas reckons it has taken about $1.6m in free tokens from one vendor this year. Another gave $65,000, and a third gave $10,000.
Both labs say they are adapting. An OpenAI spokeswoman said GPT-5.6 Sol was trained to be far more token efficient. Anthropic released a powerful lower-cost model on Friday. An executive said customers can choose intelligence or price inside its own ecosystem. Both say they support open-weight models.
Why the gap is so wide
The price difference is not charity. Electricity costs less in China, and new data centres there meet less local resistance. Chinese firms are also willing to run thin margins to win share and become the default.
Export controls may have helped. Cut off from the best Nvidia chips, Chinese labs had to squeeze more out of worse hardware. “For the money [a Chinese AI company would] spend on an Nvidia chip, they can buy 10 local chips from Huawei or other local chipmakers,” George Chen of the Asia Group told Fortune.
From Coinbase to DoorDash
Household names are doing it too. Coinbase chief executive Brian Armstrong said in June that the exchange had halved its AI spending. It did so by pushing staff towards Kimi and Z.AI’s GLM models.
DoorDash sends what its chief technology officer Andy Fang calls “lower-level work” to Kimi, for “better quality [at] cheaper cost”. Airbnb has used Alibaba’s Qwen for customer service. Cursor built its own Composer 2 coding model on Kimi foundations.
The result shows up in usage. Chinese models took 57% of the tokens US firms consumed on OpenRouter during one week in July. At one point mid-month, all five of the marketplace’s top models were Chinese.
It reaches well beyond developers. IDC surveyed 260 decision-makers at US firms with more than 1,000 staff. Of those, 47% said they used a made-in-China model for at least one use case, The Daily Upside reported. One in five reported extensive use.
The IPO problem
Here is why this lands harder than an ordinary price war. OpenAI and Anthropic spent years racing to build the most capable models. Then they found that plenty of customers want basic ones. Both are preparing to list publicly. The Journal reports the shift is threatening the valuations they carry into those listings.
It also explains the lobbying. The fight over restricting Chinese open-weight models has run all month. It stretches from the startups opposing a ban to the giants who signed a letter defending open models on Friday. The price sheet is what that argument is actually about.
The market has already voted once. Moonshot announced K3 on 16 July. Nvidia lost almost $600bn in value and briefly ceded its place as the world’s most valuable company to Apple.
Reasons to keep your head
Not everyone thinks the sky is falling. On TechCrunch’s Equity podcast, Sean O’Kane argued the industry does this every time. “We’re now a week out and I don’t think anybody’s feeling like the end is nigh like they were a week ago,” he said.
His colleague Kirsten Korosec asked the sharper question about a ban. “Are we accelerating and ensuring that Americans win the AI race, or are we ensuring that certain frontier labs do better than others?”
Real frictions remain. Gartner’s Igor Marchal warned about going public with it. Naming a Chinese model invites suspicion at best, he wrote, and reputational damage and procurement pushback at worst. Congress is probing Airbnb and Cursor over their use of such models.
The big clouds are also holding back. Amazon’s Bedrock, Microsoft’s Azure Foundry and Google’s Vertex AI still do not support these Chinese open-weight models. Bloomberg Intelligence analysts Mandeep Singh and William Tong made the point after their token-volume mix reached 68%, following the launches of K3 and GLM-5.2.
Today the weights land
Moonshot is due to publish K3’s weights for public download on Monday, Bloomberg reported. Anyone will be able to download, modify and host the model. Daily sales at the Beijing company have risen at least sixfold since K3 launched.
That date was always in the diary. When Moonshot unveiled K3 on 16 July, the weights were the missing proof, promised for 27 July. They arrive into a market that spent the intervening fortnight learning to shop around.
Sam Altman saw the mood shift coming. “This is the first year where AI spend has been a big topic,” the OpenAI chief executive told CNBC this month. “And all of a sudden, it’s a very big topic.”
