The same pile of AI output: $50 from Anthropic, 87 cents from DeepSeek
US companies have swapped tokenmaxxing for thrift-maxxing, mixing cheap Chinese models into their stacks and reserving the expensive ones for the hard parts. Chinese models took 57 percent of US token usage on OpenRouter in one July week.
Here is the arithmetic that changed corporate behaviour this month. Roughly 750,000 words of AI output cost about $50 from Anthropic’s Fable, according to Fortune’s price comparison. The same volume costs about 87 cents from DeepSeek’s V4-Pro, $4.40 from Z.AI’s GLM-5.2, and $15 from Moonshot’s Kimi K3, which counts as expensive by Chinese standards. The Wall Street Journal has a name for what companies did next: thrift-maxxing. Last year the badge of honour was tokenmaxxing, burning as many tokens as possible. Now the goal is finishing the job with the cheapest model that can do it.
A token, if the word is new to you, is roughly a chunk of a word, and it is the unit AI providers bill by. The first corporate response to rising bills was rationing, capping what employees could spend. The second is substitution, and that is the one that hurts the labs. It mostly is not wholesale replacement. Telnyx now runs 1,400 agents on Z.AI models, with Anthropic’s Fable acting as the conductor that plans the work and OpenAI’s Sol reviewing the output. The legal startup Harvey trained GLM-5.2 itself and gave it a button to call Fable 5 when a task looks genuinely hard. Coinbase said in June it had halved its AI spending by steering staff toward Kimi and GLM. Airbnb has used Alibaba’s Qwen for customer service. In one week in July, Chinese models accounted for 57 percent of the tokens US firms consumed on the marketplace OpenRouter, and an IDC survey of 260 decision-makers at large US companies found 47 percent using a Chinese-made model for at least one use case.
The price gap is not generosity. Electricity is cheaper in China, new data centers there face less local opposition, and Chinese labs are willing to run thin margins to become the default. Export controls may have helped rather than hurt: cut off from the best Nvidia chips, those labs had to extract more from worse hardware. The consequence for OpenAI and Anthropic is awkward, because both spent years racing to build the most capable model and then found that a large share of customers wanted a merely adequate one. Both are preparing to go public, and the Journal reports the shift is pressing on the valuations they carry into those listings. Anthropic’s answer on Friday was Opus 5, a cheaper high-end model. Worth keeping expectations grounded on both sides: TechCrunch’s Sean O’Kane notes the industry panics on this cycle every time, and a week after the Kimi K3 shock the mood had already cooled. Real frictions remain too. Amazon Bedrock, Microsoft Azure Foundry, and Google Vertex AI still do not carry these Chinese open-weight models.
What this means for you: If you pay for an AI subscription, this competition is why prices and limits keep moving, mostly in your favour. If you build anything with an API, the practical move is the one these companies made: stop defaulting to the most capable model for every call, and route only the genuinely hard steps to it. You can test the alternatives in an afternoon on a marketplace like OpenRouter. If you handle sensitive company data, the caution is real and not just geopolitics: security researchers warn about sending proprietary code to any foreign provider, which is exactly why open-weight models you can host yourself are part of the appeal.
Sources
Source: https://thenextweb.com/news/thrift-maxxing-chinese-models-openai-anthropic-ipo-valuations
Computer science teachers are quietly rewriting their exams because of AI
A survey of 763 educators in 49 countries found that 68 percent have already changed how they test students. Oral exams and code defenses are back, and the skill being graded is shifting from writing code to understanding it.
