硅谷AI大模型掀起价格战

OpenAI和Anthropic两大硅谷AI巨头近期大幅下调旗舰模型价格，引发业界价格竞争升级。1Anthropic于7月24日推出Claude Opus 5，定价为每百万Token输入5美元、输出25美元，较其前代产品Fable 5（输入10美元、输出50美元）降低50%。1随后OpenAI发布GPT-5.6，输入价格同样降至每百万Token 5美元，输出价格为30美元，亦为Fable 5价格的一半。1

这场价格竞争反映出市场需求的深刻变化。1企业用户的采购决策已从单纯追求性能指标转向关注单位任务成本，同时来自中国大模型的竞争压力也不容忽视。1数据显示，Kimi K3的定价为输入每百万Token约3美元、输出约15美元，价格更具竞争力。1与此同时，达到GPT-3.5水平的模型推理成本在2022年11月至2024年10月间下降了超过280倍，体现了技术进步对成本结构的深刻影响。1

尽管大幅降价，两家公司仍面临盈利压力。1OpenAI推理成本在2025年增长了四倍，导致调整后毛利率从40%降至33%，同时该公司将截至2030年的预计算力支出上调至近7500亿美元。1市场需求方面，企业用户的多模型采用趋势加剧竞争：37%的企业已同时使用五款以上模型，高于前一年的29%。16月份Anthropic的企业付费采用率达到42.4%，OpenAI为39.5%。1此外，ChatGPT的周活跃用户已突破10亿，Codex和ChatGPT Work的活跃用户合计突破1000万。1

Leading artificial intelligence companies in Silicon Valley are engaged in aggressive pricing competition for their flagship large language models. Anthropic released Claude Opus 5 on July 24, pricing it at $5 per million tokens for input and $25 per million tokens for output—half the cost of its Fable 5 predecessor while maintaining comparable performance 1. OpenAI subsequently launched GPT-5.6 with input pricing of just $5 per million tokens and output pricing of $30 per million tokens 1. This pricing offensive reflects a fundamental shift in how enterprise customers evaluate AI services, with organizations increasingly prioritizing cost-effectiveness per task over raw performance metrics alone 1.

The competitive dynamics extend beyond these two companies, with pricing pressures originating partly from emerging rivals. Kimi K3 is available at approximately $3 per million tokens for input and $15 per million tokens for output 1. Meanwhile, enterprise adoption patterns show fragmentation across vendors: 37% of enterprises now simultaneously use five or more AI models, up from 29% in the previous year 1. By June, Anthropic had achieved a 42.4% enterprise paid adoption rate compared to OpenAI's 39.5% 1.

The pricing war reflects mounting cost pressures on AI providers. OpenAI's inference expenses quadrupled in 2025, compressing adjusted gross margins from 40% to 33% 1, while the company increased its projected computational spending through 2030 to nearly $750 billion 1. However, the industry has achieved substantial efficiency gains: the inference cost to achieve GPT-3.5-level performance dropped by more than 280 times between November 2022 and October 2024 1. Despite competitive pressures on pricing, ChatGPT has reached over 1 billion weekly active users 1.