过去两个月，OpenAI、Anthropic、DeepSeek等AI公司集中启动Token价格战，通过大幅降价、免费额度赠送和套餐优惠等手段争夺用户[1]。OpenAI的举措尤为激进，宣布将GPT-5.6 Luna的API价格下调80%，Terra降价20%，分别降至每百万Token 0.2美元/1.2美元和2美元/12美元[1]。此外，OpenAI还向YC创业公司提出每家最高可获得200万美元Token的投资方案[1]。
Agent应用的高速增长正推高Token消耗。Codex每周活跃用户已超过500万，较2月推出时增长超6倍[1]。用户使用深度也在提升，其中80.6%的Codex个人用户至少提交过相当于人类工作30分钟以上的任务，25.6%提交过8小时以上的任务[1]。Claude Code也表现强劲，年化收入已超过25亿美元，较年初翻倍，企业订阅数量增长4倍[1]。微软研究显示，编程Agent持续使用者合并的PR数量提高约24%[1]。
这场补贴竞争背后是AI公司商业策略的转变。业内人士坦言用户忠诚度正在消散——Pylon的CEO表示从一家模型供应商处获得价值约160万美元的免费Token，并直言"我完全看不到任何忠诚度"[1]。在头部模型差距缩小、用户迁移成本下降的背景下，AI公司正从技术竞争转向商业化竞争，试图通过补贴争夺用户长期工作流和消费习惯[1]。
Over the past two months, major artificial intelligence companies including OpenAI, Anthropic, DeepSeek, and MiniMax have simultaneously rolled out token discount strategies in a competitive bid to attract and retain users [1]. The promotional tactics span price reductions, free token allocations, and bundled subscription packages [1].
OpenAI has made particularly aggressive moves, slashing API prices for GPT-5.6 Luna by 80% to $0.2 per million tokens and reducing Terra pricing by 20% to $2 per million tokens and $12 for higher-tier access [1]. The company has also offered a novel investment approach to Y Combinator startups, providing eligible founders up to $2 million in token credits per company [1]. This pricing warfare reflects underlying market pressures: Codex's weekly active users have surpassed 5 million, more than six times the figure recorded when the service launched in February [1]. Among Codex's individual users, 80.6% have submitted tasks equivalent to at least 30 minutes of human work, while 25.6% have submitted tasks requiring 8 hours or more [1].
Anthropic's Claude Code has demonstrated similarly robust growth, achieving annualized revenues exceeding $2.5 billion—double the figure from the beginning of the year—alongside a fourfold increase in enterprise subscriptions [1]. The competition has become so intense that Pylon's CEO acknowledged receiving approximately $1.6 million in free tokens from one model supplier, while candidly observing, "I see absolutely no loyalty whatsoever" [1]. Microsoft research has further validated the productivity impact of agent technology, finding that consistent users of programming agents increased their merged pull requests by approximately 24% [1].
