为什么开源 AI 的崛起目前还没有伤到 Anthropic

周一，Decagon 首席执行官张峥发布了一个引发思考的新理论，标题为「关于企业中的开源 AI，所有人都错了」。这篇文章探讨了当今 AI 经济中一个最有趣的矛盾：他说，越来越成熟的 AI 部署正在转向更轻量的模型，甚至在他自己的公司也是如此。但在昂贵的顶尖模型上的总支出却几乎没有变化。

这是思考前沿模型与开源模型关系的一种新方式。按照张峥的说法，它们并不是竞争对手，开源模型的成功并没有以前沿实验室的损失为代价。相反，它们是同一生命周期的两个阶段：昂贵的前沿模型被用来验证那些可以在成熟后交给更便宜的开源替代方案的用例。

随着更成熟的用例转向更轻量的模型，新的用例不断出现——而前沿模型的总支出几乎没有下降。

张峥并没有给出太多数据来支持这一观点，但数据并不难找。Vercel 的 AI 网关仪表板显示，仅在过去一周，DeepSeek 就迅速在 token 数量上冲到第一，目前处理着通过该公司基础设施的 token 中的三分之一多一点。Z.ai——流行模型 GLM-5.2 背后的实验室——在同一时期也上升到了可观的第四名。

但如果向下滚动到总的 token 支出，你会发现 Anthropic 仍然占据着该平台上 AI 总支出一半以上的份额。考虑到近期的变化很大一部分来自 Anthropic 自身上调的价格，其份额在过去一个月略有下降，但并不显著。

图片来源： Vercel 仪表板 / 数据导出

OpenRouter也讲述了一个类似的故事，它覆盖了一个更大（但企业属性稍弱）的市场领域。DeepSeek V4 Flash 在总用量上是主要赢家，每周处理 5.3 万亿 token。最受欢迎的前沿模型 Opus 4.8 处理量刚超过 2 万亿 token。OpenRouter 没有按总支出对模型进行排名，但它记录了 Opus 4.8 的平均 token 成本大约是 V4 Flash 的 23 倍（每百万 token 1.37 美元，而 V4 Flash 仅为 6 美分），这意味着 Opus 可能仍然占据了支出的大部分。

这些数据还没有涵盖最新成员——英伟达的 Nemotron，该模型凭借英伟达强大的关系网络和自身极高的可适应性，有望跃居前列。

这些数字并不完全能证明张峥关于 AI 生命周期的观点，但它们确实表明，像 Anthropic 这样的前沿实验室并没有因开源的崛起而受到太大冲击——至少目前还没有。一种解释是，AI 可处理任务的市场增长如此之快，以至于顶级模型仅通过主导早期部署阶段就能维持自己的位置。正如张峥所说：

「前沿实验室将继续掌控发现阶段。开源将逐渐掌控生产阶段。」

另一种解释可能是，即使客户转向开源，许多用例的难度如此之大，以至于无法完全用更便宜的替代方案取代。

无论如何，这种模型的双轨经济可能成为 AI 经济中一个相对稳定的特征。

早在去年 9 月，我就在写文章探讨基础模型实验室最终会「向星巴克卖咖啡豆」的可能性——即作为商品化基础设施的提供者，而应用层则收获大部分利益。这一预测的某些部分已经成真：例如，垂直 AI 企业转向了更轻量的模型，并且「GPT 套壳」创业公司的经济状况基本保持稳定。

但我们也看到，按每 token 计算，前沿模型提供商能够守住市场上最令人向往的部分——高昂的 token 定价。而这在短期内似乎不会改变。

On Monday, Decagon CEO Jesse Zhang published a provocative new theory, posted under the title “Everyone is wrong about open source AI in the enterprise.” The post grapples with one of the most interesting contradictions of today’s AI economy: More mature AI deployments are switching to lighter models, he says, even at his own company. But the overall spend on expensive state-of-the-art models has barely budged.

It’s a new way to think about the relationship between frontier and open source models. In Zhang’s telling, they aren’t competitors, and open source models’ success isn’t coming at the expense of frontier labs. Instead, they’re two phases of the same life cycle, with expensive frontier models being used to prove out use cases that can be passed along to cheaper open source alternatives as they mature.

As more mature use cases switch to lighter models, new use cases keep arising — and the overall spend on frontier models barely goes down.

Zhang doesn’t give much data to support the point, but the data isn’t hard to find. Vercel’s AI gateway dashboard shows that, in just the past week, DeepSeek has surged into the lead for token volumes, now processing just over a third of the tokens passing through the company’s infrastructure. Z.ai — the lab behind the popular GLM-5.2 model — jumped into a respectable fourth place over the same period.

But if you scroll down to overall token spend, you’ll see Anthropic still accounts for more than half of the overall AI spend on the platform. Given that much of the recent change comes from Anthropic’s own rising prices, the share has dropped slightly over the past month, but not significantly.

Image Credits: Vercel dashboard / data export

OpenRouter tells a similar story, capturing a much larger (but slightly less enterprise-y) segment of the market. DeepSeek V4 Flash is the main winner on overall usage, processing 5.3 trillion tokens weekly. The most popular frontier model, Opus 4.8, handles just over 2 trillion. OpenRouter doesn’t rank models by total spend, but it registers the average token cost for Opus 4.8 as roughly 23x higher than V4 Flash ($1.37 per million tokens, compared to just 6 cents), which would mean Opus was still probably capturing the lion’s share of spending.

Those figures don’t even capture the newest arrival, Nvidia’s Nemotron, which is poised to leap to the front of the pack by virtue of Nvidia’s strong connections and the model’s own extreme adaptability.

Those figures don’t fully prove Zhang’s point about the AI life cycles, but they do show frontier labs like Anthropic aren’t suffering too much from the rise of open source — at least not yet. One explanation is that the market of AI-addressable tasks is growing so fast that the top models are able to maintain their position just by dominating early-stage deployments. As Zhang puts it, “The frontier labs will keep owning discovery. Open source will increasingly own production.” Another explanation might be that, even as clients move to open source, many use cases are so difficult that they can’t be entirely replaced with cheaper alternatives.

Either way, this two-tiered economy of models may become a relatively stable feature of the AI economy.

As recently as last September, I was writing about the possibility that foundation labs would end up selling coffee beans to Starbucks — that is, serving as commodity inputs while the application layer reaped the benefits. Some parts of that prediction came true: Vertical AI plays switched to lighter models, for one, and the economics of “GPT wrapper” startups have remained mostly stable.

But we’re also seeing that, token for token, frontier providers have been able to hold on to the most desirable part of the marketplace — the premium token price. And that doesn’t seem likely to change any time soon.