In the first week of March token consumption by the Chinese models increased by 35\u00a0per cent, reaching 4.19\u00a0trillion tokens, whereas US models saw the use of 3.63\u00a0trillion tokens compared with the week before. Tokens are small chunks of text, words or parts of words, that AI models use to read and generate language. This data is drawn from Open Router, a widely used platform that lets developers access hundreds of AI models.\u00a0Ticking upwards like a taxi metre, the more tokens you use, the more it costs.<\/p>\n
Why are Chinese models proving more popular? They are cheaper, charging a fraction of the cost of their US peers. For a million input tokens Minimax charges $0.30 for its M2.5 model compared with $5 for Anthropic\u2019s Claude Opus 4.6 model. When it comes to output, Deepseek charges $0.28 per million tokens for its V3.2 model, compared with $15 for OpenAI\u2019s ChatGPT<\/a>-5.4 model. Ouch.\u00a0<\/p>\n For many people\u2019s purposes, the cheaper versions do the job. <\/p>\n How are they cheaper? Minimax claims that it spent just half a million dollars renting the computer power that it needed to train its M1 model last year, which if true, is nearly two hundred times cheaper than estimates of the training cost of ChatGPT-4o. Among other things, energy in China is far less of a constraint.<\/p>\n Minimax remains a minnow compared with OpenAI, which turned over $20\u00a0billion in 2025 but it is growing, with 70\u00a0per cent of Minimax\u2019s $79\u00a0million revenue last year coming from outside China (the company floated in Hong Kong in January). <\/p>\n The top three Chinese models are notably all open source, meaning the underlying models are free to download, naturally keeping the cost down. Closed models, such as Anthropic\u2019s, charge for both the model and the computing power to use them. <\/p>\n A paper from Massachusetts Institute of Technology found that \u201creallocating demand from observably dominated closed models to superior open models would reduce average prices by over 70\u00a0per cent\u201d.\u00a0<\/p>\n It seems consumers are catching on, even those you might not expect.\u00a0Brian Chesky, the Silicon Valley stalwart who founded Airbnb, said in October that his company used Qwen by China\u2019s Alibaba to power its customer service agent because it is \u201cfast and cheap\u201d.<\/p>\n Business newsletter<\/p>\n The business editor\u2019s exclusive analysis of all the latest financial and economic news.<\/p>\n \t\t\t\t\tSign up with one click<\/p>\n Then there is the Open Claw<\/a> phenomenon. Some of the push towards cheap Chinese tokens is being led by the tech world\u2019s obsession with the downloadable agent which can do online admin, take control of your inbox, make reservations or check you in for flights through messaging systems like WhatsApp.<\/p>\n Open Claw might be free to download, but the AI used to power it, is certainly not. A tool that powerful needs an awful lot of tokens, the cheaper the better.<\/p>\n Last month Peter Steinberger, Open Claw\u2019s founder, enabled support for two Chinese models as the brains behind the agent. Ironically, Steinberger, because of his rip-roaring success building this maverick machine, is now employed by OpenAI.<\/p>\n This exponential rise in Chinese AI use is happening despite constraints imposed by the US. There is an argument that American restrictions on exports of chips such as Nvidia<\/a>\u2019s have made China more innovative.\u00a0<\/p>\n What they lack in semiconductors, they make up for in engineering. <\/p>\n Technically, Chinese labs have perfected a more efficient \u201cmixture of experts\u201d (MoE) architecture, which means instead of running the whole AI brain for every question, it only uses the relevant parts.\u00a0<\/p>\n China\u2019s share of researchers with more than a thousand citations in academic articles has been rising steadily and is now on par with the US (both about 34\u00a0per cent). Academic citations are a proxy for quality and influence, they are not just about how much research someone is producing, but its importance. <\/p>\n Throw in significant Beijing government funding with a national strategy to achieve global AI leadership by 2030 and China\u2019s success is not a surprise.<\/p>\n![\"Dario](\"https:\/\/www.newsbeep.com\/uk\/wp-content\/uploads\/2026\/03\/3746d453-deaf-4f6b-957c-0c36086ab612.jpg\")
Dario Amodei, chief executive of AnthropicReuters<\/p>\n