我每天追踪大语言模型 API 的价格,发现在同一模型上存在 33 倍的成本差距。

1 分•作者: costpertoken•20 天前
我一直在寻找最新的大型语言模型(LLM)定价对比文章,但发现它们很快就过时了。因此,我构建了一个爬虫,每天都会抓取 OpenRouter 目录(包含 425 个模型,58 家提供商)的完整信息,并与前一天的数据进行比对。这个爬虫已经无人值守运行了大约一周:https://costpertoken.dev 数据分析中有一个意外的发现让我感到惊讶:对于完全相同的模型,一个用于编码代理的调用(输入长上下文,输出代码)比一个用于批量分类的调用,每次调用的成本大约高出 33 倍。我在 GPT-5.6、Claude Sonnet 5 和 Gemini 3.7 Flash 上都进行了验证,结果分别是 33.6 倍、33.3 倍和 33.3 倍——尽管这三个模型来自不同的定价体系,但结果却几乎一致。这似乎是调用模式的特性,而不是供应商的差异。相关文章在此:https://costpertoken.dev/guides/why-agent-calls-cost-more-than-chat-calls/ 该网站是一个静态的 Cloudflare Worker。数据收集通过一个本地的 launchd 作业运行,该作业独立于任何 AI 模型或 API,专门设计成不会在静默中停止工作。 欢迎提供反馈和定价修正——目前是第 8 天,还有一些不完善之处。
查看原文
I kept finding LLM pricing comparison posts that were already stale, so I built a scraper that snapshots the full OpenRouter catalog (425 models, 58 providers) every day and diffs it against the previous day. It&#x27;s been running unattended for about a week now: https:&#x2F;&#x2F;costpertoken.dev<p>One thing that fell out of the data surprised me: on the exact same model, a coding-agent-shaped call (large context in, code out) costs roughly 33x more per call than a bulk-classification-shaped call. I checked this on GPT-5.6, Claude Sonnet 5, and Gemini 3.7 Flash and got 33.6x, 33.3x, and 33.3x respectively -- almost identical despite three unrelated pricing tables. Seems to be a property of the call shape, not the vendor. Writeup here: https:&#x2F;&#x2F;costpertoken.dev&#x2F;guides&#x2F;why-agent-calls-cost-more-than-chat-calls&#x2F;<p>The site is a static Cloudflare Worker. Data collection runs via a local launchd job independent of any AI model or API, specifically so it can&#x27;t silently stop working.<p>Feedback and pricing corrections welcome -- it&#x27;s day 8, so there are rough edges.