字号·· | 护眼
商业内幕

认识4位硅谷工程负责人:他们高额的Claude账单解释了Anthropic的崛起Meet 4 Silicon Valley engineering leaders whose hefty Claude bills explain Anthropic's rise

点「原文对照」整页切到原文,或双击某段只看那段的原文。

在旧金山举行的Anthropic开发者大会上,一张海报宣传了该公司的Claude产品。A poster at Anthropic's developer conference in San Francisco advertises its Claude product.

Anthropic本周首次公开募股的目标估值为1.4万亿美元,而这一目标之所以可能实现,靠的是一些花钱极为大方的客户。Anthropic的数千家企业客户以及Claude Code的流行,正将巨额资金送入其金库。在所谓的“token最大化”热潮过去数月后,一些工程负责人对Business Insider表示,他们在单个项目上的Claude账单仍高达数万美元。

Anthropic is targeting a $1.4 trillion valuation in its initial public offering this week, and that's possible because of some very spendy customers. Anthropic's thousands of business customers and the popularity of Claude Code are funneling huge sums into its coffers. Months after the so-called "token-maxxing" craze, some engineering leaders told Business Insider they're still racking up Claude bills worth tens of thousands of dollars on individual projects.

另一些人则已经预期,旗下部分工程师偶尔在token(AI输出的计量单位)上的花费会超过他们的薪水。Anthropic的IPO招股书称,未在企业账户上使用Claude Code的客户平均每月花费不到75美元,而使用该编程工具的前5%用户平均每月花费约1500美元。总体而言,Anthropic报告称,过去一年支出超过10万美元的企业客户有11900家,支出达到1000万美元或以上的有250家。

Others have come to expect that some of their engineers will occasionally spend more on tokens, the units of AI's output, than they make in salary. Customers who didn't use Claude Code on their business accounts spent less than $75 a month on average, Anthropic's IPO prospectus said, while the top 5% of users using the coding tool averaged around $1,500 a month. All told, Anthropic reported having 11,900 business customers that'd spent $100,000 over the previous year, and 250 that'd spent $10 million or more.

Business Insider采访了几位在Anthropic上大额支出的客户:他们都是已拥抱AI编程、并在很大程度上偏爱Claude而非竞争对手的公司的工程负责人。Grafana Labs是一家可观测性初创公司,最近一次估值时为90亿美元,该公司对向AI编程工具投入大量资金毫无顾虑。其首席技术官汤姆·威尔基对Business Insider表示,公司的token账单已跃升至每年数千万美元的区间。

Business Insider spoke with several big Anthropic spenders: engineering leaders at companies that have embraced AI for coding and largely favored Claude over its competitors. Grafana Labs, an observability startup last valued at $9 billion, has no problem with pouring cash into AI coding tools. Its chief technology officer, Tom Wilkie, told Business Insider that the company's token bills have jumped into the tens of millions-a-year range.

他说,其中约三分之二花在Claude上,其余大部分流向OpenAI的Codex。“我们不担心这笔支出,”威尔基说,“至少目前还不担心要去控制它。”威尔基手下的一个团队曾在单月花费数万美元购买token,以提升Grafana一项服务的性能,结果使公司的亚马逊云服务账单减少了六位数。

It spends about two-thirds on Claude, with the rest mostly going to OpenAI's Codex, he said. "We're not worried about the spend," Wilkie said. "We're not worried about controlling it, just yet." One of Wilkie's teams spent tens of thousands of dollars on tokens in a single month to improve the performance of one of Grafana's services — resulting in a six-figure cut to the company's Amazon Web Services bill.

他还看到,一位非工程师岗位的员工把客户反馈转交给Anthropic的Fable模型,在token上花费了超过3万美元,并让一批成功的改动进入了产品。威尔基说,工程师们现在更愿意切换项目,因为编程语言的门槛感觉降低了。在一次数据分析师因让模型持续运行而意外累积了数万美元token支出之后,威尔基说,Grafana现在以24小时为期限开放最强AI模型的使用权限。

He also saw a non-engineer route customer feedback to Anthropic's Fable model, spend more than $30,000 dollars on tokens, and get a cluster of successful changes into the product. Engineers have gotten more comfortable switching projects because the coding language barrier feels lower now, Wilkie said. After a data analyst accidentally racked up tens of thousands of dollars in token spend by leaving a model running, Wilkie said Grafana now gives access to the most powerful AI models in 24-hour increments.

不过,他说,这仍然是“不问任何问题”。威尔基说:“我们确实想谨慎对待预算和模型访问控制带来的寒蝉效应,因为你根本不知道下一项重大创新会来自哪里。”布莱恩·比绍夫在Theory Ventures负责AI,领导一支由六名工程师组成的团队,为这家旧金山风投公司的投资者、人才与关系主管以及情报主管构建定制软件。

Still, it's "no questions asked," he said. "We really want to be careful with the chilling effect of budgets and access control to models because you just don't know where the next big piece of innovation is going to come from," Wilkie said. Bryan Bischof, who runs AI at Theory Ventures, leads a team of six engineers who build custom software for investors, heads of talent and relationships, and intelligence leads at the San Francisco venture firm.

他个人在一个月内累计了六位数的token支出,主要用于Claude,并说他如果花更多时间运行模拟,很容易突破100万美元。他和他的工程师有时会使用开放权重模型,这些模型远比Anthropic的高端选项便宜。比绍夫说,由于时间如此宝贵,他通常乐于让团队在高端工具上花钱。他说:“如果一个非常智能的模型能大幅缩短周期时间,那绝对值得。”

He has personally racked up six figures of token spend in one month, mostly on Claude, and said he could easily crack $1 million if he spent more time running simulations. He and his engineers sometimes use open-weight models, which are far cheaper than Anthropic's top-end options. Bischof said that since time is so valuable, he's often happy for his team to spend on premium tools. "If a very intelligent model is going to dramatically reduce the cycle time, then that's absolutely worth it," he said.

比绍夫说,他的公司也可能投资自己的常驻AI代理,类似于Meta的Muse和OpenAI的Dots。初创公司Menos AI为投资者运行一款AI驱动的研究软件,联合创始人克里斯·杨告诉《商业内幕》,其工程团队通过编程工具Cursor使用Claude Code或Claude模型。他说:“因为Claude在编程能力方面一直进化得非常好,我甚至没有太多机会或意愿去尝试Codex。”

Bischof says his firm may also invest in their own always-running AI agents, similar to Meta's Muse and OpenAI's Dots. At Menos AI, a startup running an AI-powered research software for investors, the engineering team uses Claude Code or Claude models through the coding tool Cursor, cofounder Chris Yang told Business Insider. "Because Claude has been evolving so, so well in terms of coding capabilities, I don't even have much opportunity or willingness to try out Codex," he said.

杨说,公司的投资者软件消耗了约90%的token成本,因为它会根据问题的难度将用户的询问分配给不同的模型。杨称自己是公司的“重度用户”之一,因为他每月在Claude token上花费3000至4000美元,同时运行多个AI智能体。杨说:“我们确实在Claude Code的使用上花了很多钱,但当我们把它与招聘一个人力资源员工相比时,它仍然是一个很大的优势。它仍然比人力劳动便宜得多。”

The company's investor software consumes around 90% of its token costs, Yang said, as it routes users' inquiries to different models depending on their difficulty. Yang called himself one of the company's "hero users," since he spends $3,000 to $4,000 on Claude tokens in a month running several AI agents at a time. "We do spend a lot of expenses on the Claude Code use, but when we compare it to a human resources hire, it's still a big benefit," Yang said. "It's still much cheaper than human labor."

Harness是一家估值55亿美元的软件开发初创公司,拥有500多名工程师。软件工程高级总监基兰·甘加达哈拉帕告诉《商业内幕》,在最近一周里,他的员工使用了2650亿个AI token,其中Claude Code和Cursor是他手下的工程师中最受欢迎的工具。他说:“我们确实看到它在不断改进,并承担越来越大一部分编写代码的人工工作。”

Harness, a software development startup worth $5.5 billion, has more than 500 engineers. Kiran Gangadharappa, a senior director of software engineering, told Business Insider that in a recent week, his staff used 265 billion AI tokens, with Claude Code and Cursor being the most popular tools among his engineers. "We definitely see it improving and taking a bigger and bigger chunk of the manual work of writing code," he said.

甘加达哈拉帕说,消耗token最多的是像重建现有软件架构或进行安全研究这样的巨型一次性任务。他说,使用Anthropic较便宜的模型之一Sonnet帮助公司削减了开支,为工程师设定每周125美元的软预算也同样有效。他预计公司的AI支出已达到“稳定”水平。甘加达哈拉帕说:“智能体从不疲倦。”

The biggest token-guzzlers, Gangadharappa said, are gigantic one-off tasks like rebuilding an existing software architecture or conducting security research. He said that using Sonnet, one of Anthropic's cheaper models, helped the company cut spending, as did creating a soft $125-a-week budget for engineers. He expects the company has reached a "steady" level of AI expenses. "Agents never get tired," Gangadharappa said.

阅读媒体原文 ↗