The cost of using generative artificial intelligence has fallen to a record low, driven by the spread of low-cost Chinese open-source models and price-cutting competition among major providers. The decline is raising alarms about the earnings models of developers such as OpenAI and Anthropic, as well as the prospects for recouping large-scale AI investments.
Silicon Data's LLM token spending index stood at $0.97 on the 31st of last month, CNBC reported on the 1st.
That is the lowest reading since the index was launched at the end of 2025, and less than half its peak this summer. The index tracks the prices at which tokens — the units large language models use to process text — actually trade in the market.
Falling token prices reduce costs for consumers using services such as ChatGPT, Claude and Gemini. For the companies supplying them, however, pricing power weakens. If the slide persists, consumers may come to take low prices for granted, further eroding providers' margins.
One of the main drivers of the decline is China's low-cost open-source models. Charles-Henry Monchau, chief investment officer at Syz Group, said open-source models such as Moonshot AI's Kimi K3 are undercutting the leading labs on price and pulling down prices across the entire market.
Competition has intensified further after OpenAI cut prices on two models in its GPT-5.6 line in late July and other providers rolled out dynamic pricing that adjusts rates according to demand. A drop in the cost of producing tokens has also contributed. Monchau said companies face token deflation, in which fixed costs tied to computing remain unchanged while revenue declines. He added that the performance gap between open-weight models and top-tier closed models is narrowing on a scale of months, and that the source of AI companies' competitiveness must therefore shift away from model performance itself toward distribution channels, memory and the ability to handle context.
The plunge in token prices is a burden for OpenAI and Anthropic, both of which are preparing initial public offerings. The two companies recently filed confidentially with U.S. regulators. With Nvidia, Microsoft and others having poured billions of dollars into expanding AI computing capacity, doubts are also mounting over whether such large investments can generate the returns expected of them.
Steve Hou, head of research at Silicon Data, said the recent price declines may signal that the market already has enough supply — combining cutting-edge and low-cost models — to handle most tasks. In other words, if prices fall faster than AI demand grows, both model providers' revenue and returns on AI infrastructure investment could fall short of market expectations.
On the New York Stock Exchange that day, growing concerns about the sustainability of AI investment weighed on technology shares. The Nasdaq composite fell 1.03%, and the Philadelphia Semiconductor Index dropped more than 2%.
Anthropic, meanwhile, unveiled the latest versions of its top-tier AI models, Mythos and Fable, the same day. The new models improve performance while requiring fewer tokens for the same tasks, an efficiency gain that lowers cost burdens.
The next-generation models, Claude Mythos 5.1 and Claude Fable 5.1, are frontier AI systems. Access to Mythos is currently limited because of cybersecurity threat concerns. Fable is a version of Mythos released to the public with safeguards attached.
