大语言模型Token市场正经历深度调整。由Silicon Data编制的LLM Token支出指数从5月的2.05美元/百万Token高位跌至8月的0.97美元/百万Token，首次跌破1美元大关，单月环比下跌29%，累计跌幅超过50%1。这一价格下行趋势与市场使用量增长形成鲜明对比——OpenRouter平台8月的Token使用量环比增长约47%，但对应的美元支出仅增长约7%1。
供给端的激烈竞争是推动价格下行的核心因素1。OpenAI对GPT系列模型进行了价格下调，同时Anthropic推出的Claude Opus 5价格仅为Claude Fable 5的一半1。在需求端，企业成本控制意识普遍提升，微软等大型科技公司开始收紧内部Token使用管理，逐步将默认工作负载切换至成本更低的方案以降低支出1。此外，大企业自建推理集群和开源模型的价格优势也进一步压低了市场定价1。
这轮调整对产业链各环节造成不同程度的冲击。基础大模型厂商面临显著盈利压力，毛利率可能遭到压缩，而中端通用模型产品最容易被开源模型替代1。
The large language model token market is undergoing a significant downturn characterized by soaring usage volumes coupled with collapsing prices. The LLM Token Expenditure Index, compiled by Silicon Data, fell to $0.97 per million tokens in August, marking the first time it has dropped below the $1 threshold 1. This represents a monthly decline of 29% and a cumulative fall exceeding 50% from its May peak of $2.05 per million tokens 1.
The price collapse reflects intensifying competition on the supply side, with major providers adjusting their pricing strategies. OpenAI has reduced prices for its GPT models, while Anthropic introduced Claude Opus 5 at half the price of Claude Fable 5 1. Despite these competitive pressures, demand continues to grow: usage on the OpenRouter platform surged approximately 47% month-over-month in August, though dollar expenditures increased by only about 7%, underscoring the disconnect between volume and spending 1.
Enterprise cost management initiatives are further pressuring the market. Microsoft has begun tightening internal token usage controls, progressively shifting default workloads to OpenAI's GPT-5.6 Sol to reduce expenses 1. The underlying drivers of this "high volume, low price" dynamic include escalating competition among suppliers, widespread adoption of cost governance practices by enterprises, self-built inference clusters by major technology companies, and the price advantages offered by open-source models 1. These combined pressures are compressing profit margins for foundational model manufacturers, with mid-tier general-purpose models facing particular vulnerability to replacement by open-source alternatives 1.
评论
还没有评论，欢迎留下第一条。