记账单位是一种很有用的东西。甚至在货币尚未普及之前，人们就已经试图找到标准化的方式来记录各种事物的价值。在《伊利亚特》中，用青铜或黄金打造的铠甲价值是以牛的数量来表示的（青铜铠甲值九头牛，黄金铠甲值100头牛）。兽皮、巨石和成堆的盐也曾被用作衡量标准。人工智能的世界则由它自己的单位——“token”（词元）所主导。
A UNIT OF account is a useful thing. Even before money was common, people tried to come up with standardised ways of accounting for things. In “The Iliad” the value of armour fashioned from bronze or gold was expressed as a number of oxen (nine for bronze, 100 for gold). Fur pelts, giant stones and piles of salt have all been used as measuring standards. The universe of artificial intelligence is governed by its own unit: the token.
这些词元通常是包含四个字符的文本片段，是AI模型读取或回应的内容。然而，词元的问题在于，它们不像牛或兽皮那样容易计算。首先，很难提前知道某项任务究竟会消耗多少词元。此外，不同模型提供商的收费差异巨大：中国初创公司DeepSeek推出的一款热门模型，其每词元价格仅为美国顶尖实验室Anthropic所开发的前沿模型的大约四十分之一。
These are snippets of text, typically four characters, that an AI model reads or responds with. The trouble with tokens, however, is that they do not work quite as neatly as cows or pelts did. For one thing, it is unclear in advance just how many a given task might consume. Model-makers also charge dramatically different prices: a popular model offered by DeepSeek, a Chinese startup, is about one-fourtieth the price per token of a cutting-edge one produced by Anthropic, a top American lab.
对于使用AI的企业而言，这无疑会引发挫败感。一个简单的任务，比如让模型处理发票，都可能产生巨额费用。为避免这种情况，一些公司会限制员工可使用的词元数量，但这又可能影响工作效率。这种混乱催生了一种新型商业模式：模型市场。其中最知名的便是OpenRouter，由Alex Atallah于2023年联合创立。在ChatGPT发布后，Atallah开始思考未来世界会呈现怎样的图景。
For firms using ai, this is a recipe for frustration. A simple task, such as asking a model to process invoices, could rack up a huge bill. To insure against this, some businesses cap how many tokens employees can consume. But that may hinder productivity. This muddle has birthed a new kind of business: the model marketplace. The best known of these marketplaces is OpenRouter, co-founded in 2023 by Alex Atallah. After ChatGPT was released, Mr Atallah began to wonder what kind of world lay ahead.
“当时我最担心的问题是：这会不会是一个‘赢家通吃’的市场，最终由OpenAI主导？”他说。他并不认为这种结果会成真，因此得出结论：市场将在AI行业中扮演核心角色。OpenRouter让客户可以轻松切换不同模型。如果Anthropic的Claude模型无法使用，他们可以选择OpenAI的模型；或者当某个实验室推出词元优惠时，也可以随时切换。OpenRouter对通过其平台使用的服务收取少量费用。
“My biggest question at the time then was: could this be a winner-takes-all market where OpenAI dominates,” he says. He doubted such an outcome, leading to the conclusion that marketplaces would play a central role in the AI industry. OpenRouter lets customers easily switch between models. If Anthropic’s Claude is down, they might use one from OpenAI. Or they might swap when a lab is offering a bargain on tokens. OpenRouter charges a small fee for use routed through its platform.
业务表现如此出色，以至于支付巨头Stripe在8月同意收购这家市场平台。据报道，它将斥资75亿美元收购这家成立仅三年的公司。OpenRouter并非唯一的市场平台。大型云计算提供商也运营着类似的服务。亚马逊通过Bedrock提供多种模型，并通过现有的云服务协议进行收费；谷歌则通过其Vertex AI服务提供相同的服务。但许多客户对这些服务持怀疑态度：这些云计算巨头不仅自己构建模型，还投资前沿实验室，因此很难保持中立。
Business has been so good that in August Stripe, a payments giant, agreed to buy the marketplace. It will reportedly pay $7.5bn for a firm that is barely three years old. OpenRouter is not the only model marketplace. Similar services are operated by the big cloud-computing providers. Amazon offers a range of models via Bedrock, billed through existing cloud-service agreements; Google offers the same through its Vertex AI service. But many customers view such offerings with suspicion: the cloud giants, which build their own models and invest in the frontier labs, are hardly neutral.
OpenRouter最初在开发者中广受欢迎，因为它提供了一种尝试多种不同模型的方式，并帮助用户快速评估哪种模型表现最佳。在ChatGPT发布后，许多AI实验室开始“基准最大化”——即打造能在特定测试（如SAT考试）中取得优异成绩的模型，但这些模型在处理更实用的任务时却表现不佳。在OpenRouter的市场平台上，客户用真金白银投票，从而形成了一套更公平的排名体系。
OpenRouter first gained a following among developers as it offered a way to play around with lots of different models and a means for quickly gauging which would perform best. After the release of ChatGPT, many AI labs took to “benchmaxxing”—making models that could pass certain tests, such as the sat, with flying colours, but struggled with more useful tasks. On OpenRouter’s marketplace customers vote with their dollars, creating a fairer ranking system.
近期，企业也开始将OpenRouter用作控制成本的工具。值得注意的是，DeepSeek以及另外两家中国模型开发商Z.ai和小米提供的廉价系统，在该平台上占据了最大的token消耗量（尽管目前OpenAI的某个模型也表现不俗）。即便如此，阿塔拉先生认为，许多企业在token使用方面仍然缺乏成熟的方法。他建议企业客户根据任务的开放性程度，将其分为低、中、高“熵值”三类。
More recently companies have turned to OpenRouter as a tool for controlling costs. Tellingly, it is the cheap systems offered by DeepSeek and two other Chinese model-makers, Z.ai and Xiaomi, that dominate token consumption on the platform (though an OpenAI model is also currently doing well).Even so, Mr Atallah argues that many businesses remain unsophisticated in their approach to token use. He recommends to his enterprise clients that they divide up tasks into low, medium and high “entropy”, based on how open-ended the work is.
对于低熵值任务（例如决定由哪个模型审查代码），廉价模型通常就足够了；许多中等熵值任务（例如实际审查代码）同样可以使用这类模型。而前沿模型则应保留用于高熵值任务，例如从头编写代码。
“我看到许多公司常犯的错误，就是没有对这些问题进行充分的分解。”
Cheap models are often sufficient for low-entropy tasks, such as deciding which model should review some code, and plenty of medium-entropy tasks as well, such as actually reviewing that code. Frontier models can be reserved for high-entropy tasks, like writing the code from scratch.“The mistake I see a lot of companies make is they just don’t decompose all of these things. It is all the same to them,” says Mr Atallah.
“对他们来说，都无所谓，”阿塔拉先生说。“这纯粹是巨大的金钱浪费。”要追踪正在塑造商业、工业和科技趋势的动向，请订阅我们的全球商业每周付费新闻通讯《The Bottom Line》。
“That’s just a massive waste of money.”To track the trends shaping commerce, industry and technology, sign up to “The Bottom Line”, our weeklyr-only newsletter on global business.