3 People with 100 AI Programmers, Burning Through $1.3 Million a Month! OpenAI: I'll Foot the Bill

marsbit发布于2026-05-17更新于2026-05-17

文章摘要

In a striking demonstration of AI-powered development, Peter Steinberger (creator of OpenClaw) shared that his three-person team spent $1.3 million in one month to run approximately 100 AI agents (primarily Codex instances). OpenAI covered the cost. The expenditure consumed 6.03 trillion tokens across 7.6 million requests. Steinberger argues that, with "fast mode" disabled, the cost falls below that of a single engineer while providing significantly greater output. This "cloud programmer army" handles core but tedious software engineering tasks: reviewing pull requests, finding security vulnerabilities, deduplicating issues, fixing bugs, monitoring benchmarks, and even generating PRs after meetings. This shifts AI's role from merely writing code to maintaining the entire collaborative fabric of a project. Steinberger's tool, CodexBar (a macOS menu bar app), tracks usage and costs across various AI coding services, highlighting how token consumption is becoming a key metric—a new "means of production." The experiment poses a profound question: if token cost ceases to be a barrier, how will software development transform? As model prices fall, the capability for small teams to leverage large numbers of AI agents could become commonplace, fundamentally altering the scale and speed of development. The future, Steinberger suggests, is arriving rapidly.

Peter Steinberger

Editor: Solomon

[New Zhiyuan Report] 3 people, 100 AI agents, burning through $1.3 million a month — The father of OpenClaw has turned software development into an "AI assembly line," with OpenAI picking up the tab.

While others show off their pay stubs, he shows off the bill — $1.3 million a month!

That's nearly 9 million RMB per month. It's left netizens utterly dumbfounded.

OpenClaw father Peter Steinberger casually posted a screenshot on X.

Peter Steinberger

But the numbers on the screenshot were anything but casual:

30-day spend: $1,305,088.81. Consumed 603 billion tokens. Made 7.6 million requests.

You read that right, 1.3 million U.S. dollars. And it's not a quarterly AI budget for some big tech company — it's the monthly usage of a three-person team.

Even more explosive: OpenAI is reimbursing this cost.

The comment section instantly went wild.

Some were amazed, some skeptical, some whipped out their calculators to figure out "how many programmers this equals."

Steinberger himself calmly responded: "With fast mode off, my cost is less than an engineer, and it really helps a lot more."

Translation: — It's genuinely cost-effective!

Other netizens were shocked by the $400k/month engineer — "The San Francisco job market is insane."

Netizen comment

Netizen comment

Others were curious about where this massive token usage went.

Peter responded that most was used for OpenClaw development.

Netizen comment

A Cloud-Based Programmer Army

The most outrageous thing is that Pete's small team only has 3 people.

They have about 100 Codex instances running long-term in the cloud, handling the dirtiest, most grueling, most mind-numbing work in software engineering —

Reviewing PRs, finding security vulnerabilities, deduplicating issues, fixing bugs, monitoring benchmarks, posting to Discord upon discovering regressions, even opening PRs directly after listening to meetings.

Thus, AI isn't just "helping you write code," but is infiltrating every crevice of software collaboration.

This is terrifying.

Because what's truly expensive in software development is communication, comprehension, context switching, review, regression, fixes, waiting, and repetitive tasks.

In the past, a team spent a huge amount of time each day on these things that don't seem like "creation" but without which the project would rot.

Now, Peter has tossed all these processes to a bunch of AI agents at once.

This is AI starting to maintain the nervous system of an organization for you.

Illustration

There's another important detail in this screenshot: it's not the OpenAI backend, but CodexBar made by Peter.

CodexBar is a macOS menu bar tool for tracking usage windows, credits, costs, and reset times for various AI programming tools.

It supports a bunch of services like Codex, Claude, Cursor, Gemini, Copilot, etc.

What used to be in a programmer's menu bar? CPU, memory, battery, network speed.

Now there's one more thing: tokens. Tokens are becoming a new "means of production."

CodexBar

A Final Word

$1.3 million a month, 3 people, 100 AI agents.

Ponder this set of numbers — three living humans, leading a hundred digital employees who don't eat, sleep, or demand raises, doing the work of an entire engineering team.

Some felt invigorated after reading this: AI finally isn't just a decorative vase for chatting! Others felt a chill down their spine: Wait, so what do we coders do in the future?

But honestly, what keeps me up at night is Steinberger's casual remark: "I'm exploring what software development would look like if token cost wasn't an issue."

Peter Steinberger

Everyone, he said "if."

The problem is, this "if" is visibly turning into "when" at an astonishing speed.

The work that costs $1.3 million today, after one price cut for models, becomes $130k. Another cut, $13k.

On that day, having 100 AI agents working simultaneously is no longer a game exclusive to Silicon Valley big shots, but a basic operation for any three-person startup team.

Three young people in a garage, holding a hundred tireless AI programmers — this image, just thinking about it is absurd.

Peter Steinberger has revealed the bottom card.

On the card it says: The future is already knocking, and it doesn't plan to wait for you to be ready.

References:

https://the-decoder.com/for-1-3-million-a-month-openclaw-founder-peter-steinberger-runs-100-ai-agents-that-code-review-prs-and-find-bugs/

https://x.com/steipete/status/2055346265869721905

https://developers.openai.com/codex/speed

This article comes from the WeChat public account "New Zhiyuan", author: New Zhiyuan

相关问答

QHow much money did the three-person team spend on AI development in one month, and who covered the cost?

AThe three-person team spent $1,305,088.81 (approximately 1.3 million USD) in one month on AI development, and the cost was covered by OpenAI.

QWhat is the name of the tool Peter Steinberger created to track AI development costs and usage?

APeter Steinberger created a tool called CodexBar, a macOS menu bar application that tracks usage windows, credits, costs, and reset times for various AI programming tools like Codex, Claude, Cursor, Gemini, and Copilot.

QApproximately how many AI agents (instances) does Peter Steinberger's team run for development tasks?

APeter Steinberger's team runs approximately 100 AI Codex instances to perform various development tasks.

QWhat kinds of software development tasks do the AI agents in the article handle?

AThe AI agents handle tasks such as reviewing pull requests (PRs), finding security vulnerabilities, deduplicating issues, fixing bugs, monitoring benchmarks, reporting regressions to Discord, and even opening PRs after listening to meetings.

QWhat is the core implication Peter Steinberger highlights regarding the future of software development with AI?

APeter Steinberger highlights that the core implication is exploring what software development would look like if token cost were not a limiting factor, suggesting a future where small teams can leverage large numbers of AI agents as a standard practice, dramatically changing the development landscape.

你可能也喜欢

一周省下3亿Token,Anthropic工程师的Claude Code缓存指南

这篇文章分享了Anthropic工程师通过有效利用Claude Code的Prompt缓存机制,一周内节省超过3亿Token的经验。核心观点是,缓存的Token成本仅为普通输入的10%,因此提高缓存复用率能显著降低使用成本并延长会话额度。 缓存机制依赖“前缀匹配”,分为系统层、项目层和对话层三层。只要会话前缀与已缓存内容一致,Claude即可直接复用,无需重新处理。Claude Code订阅版的缓存有效时间(TTL)为1小时,而API默认为5分钟。 文章强调了几个关键的使用习惯以避免“打断”缓存: 1. 不要让会话空置超过1小时,否则缓存过期,重新开始成本更低。 2. 切换任务时,建议进行清晰的会话交接(如使用自定义的handoff技能总结进度后新建会话),而非使用可能破坏缓存的`/compact`命令。 3. 避免在会话中途切换模型(包括开启“Opus plan”模式),因为不同模型使用独立缓存,切换会导致缓存失效。 4. 大文档应放入Projects中管理,而非反复粘贴进对话,以获得更好的缓存优化。 此外,文章提到Anthropic内部会监控缓存命中率,因为这直接影响服务成本和用户体验。通过理解并遵循这些简单的原则,用户可以在不改变使用习惯的情况下,更高效地利用Claude Code,节省Token消耗。

marsbit31分钟前

一周省下3亿Token,Anthropic工程师的Claude Code缓存指南

marsbit31分钟前

链上数据暗示,比特币底部可能还需两个月才会到来

过去一周,比特币在7万美元高位区间横盘整理,未能收复自5月中旬以来市场多头屡攻不下的关键心理价位82,000美元。值得注意的是,76,000美元价位已连续三周经受考验并成功守住,形成了一个实质支撑区。然而,一项链上指标可能正在发出比特币历史上最清晰的底部信号。 关键看跌信号源于投资者成本基础数据。5月22日,分析师CryptoChan分享了一个历史上可靠的底部指标数据,该指标基于两个已实现价格带构建:代表长期持有者平均成本的6个月至10年已实现价格(目前为60,316美元),以及代表更广泛市场平均成本基础的0至10年已实现价格(目前为64,412美元)。这两个价格带之间的比率反映了长期持有者相对于整体市场所承受的压力程度。历史上,当该比率跌破0.936并随后回升至1.0时,每次都精准标记了比特币前几个周期的底部时刻。当比率触及1.0,意味着即使是最坚定的长期持有者也处于亏损状态,此时抛压已完全耗尽,市场情绪陷入极度恐慌。 历史数据显示,在2015年熊市底部,该比率从0.936回升至1.0用了59天;在2018-2019年熊市底部用了66天;在2022年11月FTX崩溃引发的底部则用了50天。目前该比率再次位于0.936。如果当前读数保持不变且历史数据重演,比特币的明确底部窗口可能会在2026年7月中下旬左右开启。 截至撰写本文时,比特币价格为75,269美元,过去一周下跌2.84%,周线和月线图也分别下跌4.65%和3.55。市场恐惧与贪婪指数为28,显示恐惧情绪显著影响市场。不过,有分析师预测未来五天内可能因空头挤压反弹至83,354美元,一个月后目标为77,741美元,而三个月目标价看至90,529美元,暗示潜在16%的涨幅。

bitcoinist4小时前

链上数据暗示,比特币底部可能还需两个月才会到来

bitcoinist4小时前

交易

现货
合约

热门文章

如何购买BILL

欢迎来到HTX.com!我们已经让购买Billions Network(BILL)变得简单而便捷。跟随我们的逐步指南,放心开始您的加密货币之旅。第一步:创建您的HTX账户使用您的电子邮件、手机号码注册一个免费账户在HTX上。体验无忧的注册过程并解锁所有平台功能。立即注册第二步:前往买币页面,选择您的支付方式信用卡/借记卡购买:使用您的Visa或Mastercard即时购买Billions Network(BILL)。余额购买:使用您HTX账户余额中的资金进行无缝交易。第三方购买:探索诸如Google Pay或Apple Pay等流行支付方法以增加便利性。C2C购买:在HTX平台上直接与其他用户交易。HTX场外交易台(OTC)购买:为大量交易者提供个性化服务和竞争性汇率。第三步:存储您的Billions Network(BILL)购买完您的Billions Network(BILL)后,将其存储在您的HTX账户钱包中。您也可以通过区块链转账将其发送到其他地方或者用于交易其他加密货币。第四步:交易Billions Network(BILL)在HTX的现货市场轻松交易Billions Network(BILL)。访问您的账户,选择您的交易对,执行您的交易,并实时监控。HTX为初学者和经验丰富的交易者提供了友好的用户体验。

335人学过发布于 2026.05.07更新于 2026.05.11

如何购买BILL

相关讨论

欢迎来到HTX社区。在这里,您可以了解最新的平台发展动态并获得专业的市场意见。以下是用户对BILL(BILL)币价的意见。

活动图片