OpenAI、Anthropic 等美国领先 AI 实验室正在发布更便宜的模型，以留住对价格敏感、正转向中国竞争对手低价替代方案的客户。
这场价格战爆发之际，不断攀升的 AI 账单正促使企业限制使用量并寻求更便宜的模型，这也帮助月之暗面（Moonshot）、DeepSeek 等中国开发者从硅谷到欧洲一路攻城略地，赢得用户。
OpenAI 近日表示，将把 GPT-5.6 Luna（其“最快且最实惠的模型”）的价格下调 80%。Anthropic 则推出了 Claude Opus 5，宣称该系统具备“前沿智能……而价格仅为该公司最强模型 Fable 5 的一半”。
据 Silicon Data 的 token 价格指数显示，这些举措已使客户为美国领先实验室模型支付的价格自 7 月中旬以来下降近四分之一。Token 是语言模型处理的数据单位，也是许多客户账单的计算依据。
这些降价标志着美国 AI 公司的一次转变——它们此前一直以专有“闭源”模型在性能上激烈竞争。而能力日益增强的中国“开源”模型——开发者可自由下载并修改——也对价格形成了下行压力。
与此同时，OpenAI 和 Anthropic 正筹划以万亿美元估值进行首次公开募股，而投资者正在寻找证据，证明该行业在 AI 上的巨额投入能够带来回报。
随着 Anthropic 和 OpenAI 将部分企业客户从固定订阅制转向按用量计费（即企业根据其消耗的计算资源付费），企业级 AI 用户正面临成本压力。
面对账单上涨，一些企业选择对 AI 使用设置上限，或测试更便宜的替代方案。DoorDash 和 Airbnb 等公司已表示，它们已开始使用中国制造的模型以控制成本。
这一转变恰逢中国实验室密集发布新模型，缩小了与领先美国模型之间的性能差距，引发美国科技行业的担忧——即便美国开发商投入巨资维持技术优势，仍可能流失客户。
AI 实验室提供一系列能力和价格各异的模型，成本还会因模型版本和所使用的“努力程度”设置而进一步变化。客户通常按输入 token（用于衡量输入模型的数据量）和输出 token（用于衡量模型生成的回复量）分别计费。
美国实验室的最新降价适用于中端产品，使其与中国产品相比更具竞争力。例如，OpenAI 将 GPT-5.6 Luna 的价格从每百万输入 token 1 美元降至 0.20 美元，从每百万输出 token 6 美元降至 1.20 美元。
Anthropic 以每百万输入 token 5 美元、每百万输出 token 25 美元的价格推出 Opus 5——为其 Fable 5 模型价格的一半。本周，该公司取消了原定于 9 月生效的 Sonnet 5 模型涨价计划。
不过，标称 token 价格并不能直接用来比较不同 AI 模型。能力更强的模型有时可以用更少的 token 或更少的尝试次数完成任务，这意味着一个按标称 token 价格看似更贵的模型，最终完成任务的成本反而可能更低。
此外，大多数模型还可以在不同的“努力程度”设置下运行，这会改变回答问题所用的计算资源，并可能同时影响性能和完成任务的最终成本。
Artificial Analysis 对模型在数学、科学、编程和推理等多个领域进行基准测试。该机构发现，Anthropic 的 Opus 5 在“中等”努力程度下，其性能和单任务成本与月之暗面（Moonshot）的 Kimi K3 在“最高”努力程度下的表现相当。OpenAI 的 GPT-5.6 Luna 在“最高”努力程度下的表现与 DeepSeek 的 V4 Flash 在“最高”努力程度下相近，但单任务成本略低于后者的两倍。
Anthropic 和 OpenAI 拒绝置评。
一位接近 Anthropic 的人士表示，Opus 5 的定价低于其旗舰产品 Fable 5，是因为这家初创公司的“模型家族就是这样构建的，与竞争对手无关。”
Mantas Lukauskas 是网站托管服务商 Hostinger 的 AI 技术负责人，该公司自 2020 年起就开始使用大语言模型。他指出，最顶尖模型的价格“持平甚至有所上涨”。
他还补充说，近期的定价变化是对 Anthropic 和 OpenAI 等公司能否守住其最先进产品价格的“第一次真正考验”：“美国实验室已经砍掉了中端产品，正在死守高端市场。”
Leading US AI labs such as OpenAI and Anthropic are releasing cheaper models as they fight to retain cost-conscious customers who are switching to cut-price alternatives from Chinese rivals.
The price war comes as rising AI bills push companies to curb usage and seek cheaper models, helping Chinese developers including Moonshot and DeepSeek make inroads with users from Silicon Valley to Europe.
OpenAI recently said that it was slashing prices for GPT-5.6 Luna, its “fastest and most affordable model”, by 80 percent. Anthropic has launched Claude Opus 5, touting the system’s “frontier intelligence… at half the price” of Fable 5, the company’s most capable model.
The moves have helped decrease prices that customers are paying for models from leading US labs by almost a quarter since mid-July, according to Silicon Data’s token price index. Tokens are the units of data processed by language models and are used to calculate many customers’ bills.
The cuts mark a shift for US AI groups that make proprietary “closed” models that have, until now, competed heavily on performance. Increasingly capable “open” Chinese models—which can be freely downloaded and tweaked by developers—have contributed to pressure on prices.
The moves also come as OpenAI and Anthropic plot initial public offerings at trillion-dollar valuations while investors seek evidence that the industry’s vast spending on AI can generate returns.
Corporate AI users face cost pressures as Anthropic and OpenAI shift some enterprise customers away from flat subscriptions and toward usage-based billing, under which companies pay according to the computational resources they consume.
Some businesses have responded to a rise in bills by imposing caps on AI usage or testing cheaper alternatives. Companies such as DoorDash and Airbnb have said they have started to use Chinese-made models in an effort to rein in bills.
That shift has coincided with a flurry of releases from Chinese labs that have narrowed the performance gap with leading US models, raising concerns in the US tech industry that American developers could lose customers even as they spend heavily to maintain their technological edge.
AI labs offer a range of models with different capabilities and prices, with costs varying further according to the version of a model and the “effort” settings used. Customers are typically charged for input tokens, used to measure data fed into a model, and output tokens, which measure what it generates in response.
The latest price cuts from US labs apply to mid-tier products and make them more competitive with Chinese offerings. OpenAI, for example, cut the price of GPT-5.6 Luna from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens.
Anthropic launched Opus 5 at $5 per million input tokens and $25 per million output tokens—half the price of its Fable 5 model. This week, the company called off a planned rise in prices for its Sonnet 5 model, which had been due to take effect from September.
Headline token prices do not provide a straightforward comparison between AI models, however. More capable models can sometimes complete a task using fewer tokens or with fewer attempts, meaning a model that appears more expensive based on the headline price of tokens can ultimately cost less.
Additionally, most can operate at different “effort” settings, which vary the computing power used to answer a question and can affect both performance and the ultimate cost of completing a task.
Artificial Analysis, which benchmarks models across areas including math, science, coding, and reasoning, found Anthropic’s Opus 5 at “medium” effort delivered similar performance and cost per task to Moonshot’s Kimi K3 at “max” effort. OpenAI’s GPT-5.6 Luna at “max” effort performed similarly to DeepSeek’s V4 Flash at “max,” but cost just under twice as much per task.
Anthropic and OpenAI declined to comment.
A person close to Anthropic said Opus 5’s pricing below its flagship Fable 5 was how the startup’s “family of models is built, so there’s no connection to competitors.”
Mantas Lukauskas, AI tech lead at Hostinger, a website hosting provider that has used large language models since 2020, noted that prices for the very best models were “flat to rising.”
He added that the recent pricing changes are the “first real test” of whether groups such as Anthropic and OpenAI can protect the cost of their most advanced offerings: “The US labs have cut the middle and are defending the top.”