Unmasking Claude Code’s No. 1 Spender: How an AI Entrepreneur Pushes the Tool to Its Limits | A Conversation with 刘小排

Summary

• The No. 1 spender’s secret is not gaming the system but a generational gap in how the tool is used. 刘小排 claimed the account named in Anthropic’s announcement—the one that burned through roughly $50K and 7.7B tokens on the $200 monthly plan. He treats Claude Code as the engine of an entire software company, covering coding, technical research, algorithmic problem-solving, and operations: “Many people think Claude Code is for programming… They still haven’t realized it can do much more than programming in the narrow sense.” His core method is to manually refine 30-60 examples until the AI’s judgment matches his own, then batch-run 1,000 similar tasks: “I gave it 900 styles and went to sleep; it was done when I woke up.”
• He is a living case study in compute demand. Asked whether this made him want to buy Nvidia, he said that if AI moves further into daily life, each person’s token consumption could reach 10x, 100x, or even 1,000x today’s level. Anthropic’s usage cap refreshes every 5 hours: “Maybe only 5% of people discover that the 5-hour limit can actually be exhausted.” The limit was tightened once on July 28 and will be tightened again on August 28.
• First-hand product feel is alpha. On May 6, the day Gemini 2.5 Pro 0506 launched—before even changing its version number, “it was really 032-something”—he tested it and concluded, “This model is simply the strongest in the world; other people just don’t know it yet.” He bought Google at just over $130 and watched it rise above $200 a few months later. “Other investors don’t write code, so they have no way to judge what makes the model good.” He has since mostly liquidated his stock holdings: “Making money trading stocks doesn’t feel as good as building products.”
• The real division of labor among models is clear. Claude has little profile on coding leaderboards, but programmers love it because more than 90% of everyday coding involves simple tasks like interfaces, logins, and payments—exactly where Claude Code is strongest. Competitive algorithms, mathematics, and research remain GPT’s strength, including o3, o4-mini, and GPT-5. Gemini plays the “architect,” with roughly 1T tokens of context versus Claude’s 200K: it produces the technical plan, which Claude Code then executes.
• The leading AI products don’t make money; the money is lower down the market. Top-tier players privately say AI products are unprofitable, but that runs against his experience. The ones losing money are the funded leaders whose “users were created through PR,” such as the six startups that raised billions; below the top tier, there are no marketing costs, labor costs are low, and founders know where the users are from day one. “Yes, I must be profitable from day one.” His 6-7-person team generates enough revenue to support “30-50 people at market rates,” and he has no interest in fundraising: “Even if you gave me money, I wouldn’t know how to spend it.”
• Ten years at Cheetah Mobile distilled a methodology. 傅盛 taught him simplicity—subtraction, not addition—turning Clean Master’s one-button interface into a product with 600M MAU; with roughly 2.1B Android devices worldwide at the time, “one in every 3 or 4 Android phones had its product.” The broader framework was “predict, break through with a single point, and go all in.” At 19, he discovered the secret that still guides him: “Technology is two words”—science and technology are separate, and he focuses on engineering, turning open-source models into commercial versions with far better inference performance.
• The mindset for builders is: “It is not for you, it is you.” Claude Code is not a tool created for you; it is an extension of yourself. His only advice is to first build an AI-enabled business that can support you: “Life is the wilderness. You need a mine before you can roam free.”

Deep dive

1. The $200 Plan That Burned $50K: The No. 1 Spender Emerges

• The story began when Anthropic engineers found an anomalous account in the backend consuming compute 7×24 without interruption on a $200/month plan. In one month it used $50K worth of model capacity and 7.7B tokens, forcing Anthropic to issue an announcement and revise its global usage limits. Two weeks later, Chinese user 刘小排 claimed it on X: “The person Anthropic was chasing, who was consuming tens of thousands of dollars a month, was me all along.”
• Before the claim, his Twitter account had just 60 followers; afterward, it grew to roughly 5,500. He admitted that he did not realize he was No. 1 until he saw the leaderboard. It was not maintained by Anthropic, but built by a leading programmer: users logged into Twitter to sync their local usage records. “I synced it and only then realized—oh, turns out they’re not that good.”
• His background: 38 years old, born in 1987, with a bachelor’s degree from Chongqing University in 2009. He spent 10 years as a product director at Cheetah Mobile before leaving on December 31, 2024, to start a company. The team has 6-7 people and no funding; at market rates, its revenue “should be enough to support 30-50 people.”

2. The Real Source of the Usage: Everything Involved in Running a Company

• 刘小排’s own description is straightforward: “I’m just an ordinary product developer. I use it for everything involved in developing overseas products for my company.” Coding is only one part; the rest includes technical research, algorithmic problem-solving, and operations. What is a software business? “You use your ideas to build a software product and make money from users around the world, primarily in developed countries.”
• Usage ramped gradually. In the first 2 days after subscribing to the $200 plan, he found he could consume $300 a day. “I paid it back in one day. That’s when I realized this product was almost too good a deal.” Before that, his main tools were Cursor, which was fast, and Augment Code, which was slower but stronger; he used GPT o3 for difficult algorithms. After starting Claude Code in May, he “gradually realized Cursor wasn’t really cutting it,” initially treating Claude Code as “a more powerful Cursor.”
• The key conceptual leap was sitting in the official documentation. Running in the background 7×24 was already a built-in feature. “If you think one step further, you realize that running 7×24 obviously isn’t about writing code—it must be executing some other kind of task. Once you see that, you shouldn’t treat it as a tool that only writes code. You should treat it as your own extension.”

3. 齐白石 and 900 Styles: Sleeping Through 1,000 Executions

• His benchmark example involved turning 齐白石 into a selectable style for an AI image-generation product. The manual workflow meant researching the artist’s life, studying a large body of work to develop an intuitive feel, building multiple prompt templates, testing images, then generating in bulk and selecting one as the icon. Done carefully, it took an hour; done roughly, at least 20 minutes. He had 900 styles. “I completed the entire thing with Claude Code, so I made 900 styles while I slept.”
• The underlying mechanism is essentially the execution of the same task 1,000 times, with multiple subtasks inside each one, including generating 9 images and selecting the best one multimodally. He manually interrupted and calibrated the first 30-60 tasks: “If you’re unsure, interrupt it. I’ll choose and tell you why.” By task 60, “its judgment was already as accurate as mine, so I could let it run.” One task took 5 minutes; 1,000 tasks took 5,000 minutes, while he slept.

4. Why Not Manus or ChatGPT: Finite Sets Versus Infinite Levels

• The key difference is controllability. 刘小排 likes Manus, but “the things it can do are limited.” It also runs on Claude underneath, but with a toolset preconfigured by its developers: “Its tools are a finite level.” Claude Code, by contrast, is “an infinite level” for him: “Whatever tool it needs, I can give it one”—including handing it the API key for an image-generation service.

5. Fast 3D: Claude Code Joins the Model Hunt

• His Fast 3D product uses a proprietary model, and he handed the entire research workflow to Claude Code. He gave it the API for opening servers and asked it to “spin up more than 10 GPU machines, deploy more than 10 well-known open-source 3D models, automatically prepare the test set, and record performance and runtime.” Without Claude Code, “the model research alone would have taken me half a month or longer.”
• Once 3 models reached the final round, he fed Claude Code their technical reports, pretraining checkpoints, and inference code, then asked: “Is there a way to greatly improve inference speed while reducing performance only slightly?” “If you don’t give it those materials, it can only make things up. But if you give it the open-source team’s technical report, what it says is relatively accurate.”
• The host pointed out what had changed. This used to be the work of a small team combining programmers, operations, and product managers. “Now you just add one Claude Code and get it done.” Domestic tech giants do the same work through collaboration between research scientists and engineering teams, “at a fairly high cost—at least higher than mine.”

6. The “Pursuit” Timeline and the Compliance Debate: Almost Unlimited, Not Unlimited

• The timeline began with Anthropic’s July 28 post. Its opening was actually positive: “This use is remarkable… we want to enable them.” The company was trying to stop abuse such as account sharing and resale. “I obviously belong to the former category. I was their biggest fan.” His efficiency peaked during the week of August 1-7, at roughly $3,000 a day versus just over $1,000 normally. He only learned that he was No. 1 when the leaderboard appeared in mid-August.
• His compliance view is blunt: “Anthropic must believe I’m compliant. Otherwise they would have banned my account already.” The actual rule is a conversation limit that refreshes every 5 hours. “Most people can’t exhaust it within 5 hours. Maybe only 5% discover that the limit can actually be used up.” Before July 28, the restriction was weaker. “July was a little more fun. It wasn’t truly unlimited; it was nearly unlimited.” Another tightening is scheduled for August 28.

7. Noise, Leaving the Leaderboard, and “Being Past the Age of Caring Who Likes Me”

• The new followers split into 2 groups. Claude Code users came to learn: “I occasionally see people in technical groups asking how to use Claude Code. They say, just follow 刘小排’s public account. I always find that moving.” Non-users made moral judgments, ranging from “You exploited capitalism’s loopholes, then came back to China to pay salaries and social security—you’re a patriot” to “You went to an all-you-can-eat restaurant and ate the owner out of business.” His response: “It’s all noise to me. To quote Trump, I’m past the age of caring who likes me and who doesn’t.”
• After a major domestic media outlet compiled more of his old material, the noise in the comments grew louder. “The first thing I did after seeing it was leave the leaderboard and delete the account. Look at it now—it definitely isn’t me.” He did not use the episode for marketing, and Anthropic never contacted him.

8. Ten “I Am” Statements: A Spiritual Idol for Working People

• His self-introduction challenge produced a string of fragments: AI entrepreneur, product manager, programmer (“Sometimes I say I can’t code, but that may be a modest way of putting it”), comedian—fans objected to his luxury-car-and-sunglasses avatar, so “today I switched to one of me riding a shared bike”—and a 得到 score above 970. “A 得到 product manager secretly told me I’m one point higher than 罗振宇.” During the pandemic, he taught himself Berkeley courses and became a certified musician who handles lyrics, composition, arrangement, and vocals himself. “It’s not very good.”
• The label that moved him most was “a spiritual idol for many working people.” These “die-hard wage workers” saw him “living the life they imagined,” became angrier than he was when his products were copied, and rewarded a public-account post summarizing his first half-year of entrepreneurship with RMB4K in tips. “It’s as if you’re living another kind of life on their behalf.” He encourages them: “Why not start with a side business?”

9. The Boundary: If You Can Define an SOP, It Can Be Automated; Sparse Rewards Cannot

• He draws 2 lines around Claude Code’s capabilities. First, it cannot handle things in the physical world. “As long as you’re in the virtual world—the world accessible through a phone or computer—if you dare to define an SOP for something, it can definitely be automated.”
• Second, it cannot handle areas with sparse rewards. HR and administration at medium and large companies are difficult to replace because performance lies in subtle distinctions, and every boss prefers something different. It can handle tasks where the world agrees on right and wrong: “If you solve the math problem correctly, it’s correct; if you write the code correctly, it’s correct. Everywhere else, nobody knows where the reward signal is. You need a person.”

10. The Secret He Discovered at 19: Technology Is Really Two Words

• At 19, during his junior year, he interned at Microsoft Research Asia in the building on Zhichun Road said to have “the highest IQ density in the world.” There he discovered “the secret I’ve benefited from to this day”: technology is 2 words. The third floor’s research institute did science; the second floor’s engineering institute did technology. “The people at the research institute looked down on the engineering institute. They thought you went into engineering because your IQ was lower.” In the innovation engineering group, he read the papers of leading researchers and turned them into working engineering.
• The insight maps directly onto his current strategy. Open-source labs do the science and publish papers. “Because I’ve always known that scientists and engineers are separate, when I read a technical report I focus only on where inference performance is weak and where the code is redundant. Tweak it a bit, and you get something of your own with much better inference performance and only slightly worse real-world results.” “It’s been nearly 20 years, and it still serves me today.”

11. 傅盛 and Cheetah: Simplicity Means Subtraction, Then Predict, Break Through, and Go All In

• In 2014, a small product he had built was acquired by 傅盛 and folded into Cheetah Mobile. The acquisition email arrived on December 31, 2014, so he waited until December 31, 2024—exactly 10 years—to leave. He still remembers the onboarding conversation: 傅盛 said, “You’re a bright, flexible young guy. Starting tomorrow, you’re a product manager.” He asked, “What is a product?” The answer: “You’ll know once you get here.” He later led a team that built products with DAU in the tens of millions.
• Cheetah’s first lesson was simplicity. During the “mass entrepreneurship” era of 2014, everyone was adding features. “Brother, don’t tell me you’re simultaneously building WeChat, Alipay, and Meituan all rolled into one to raise money?” Cheetah believed in subtraction. Clean Master had one big button and reached 600M MAU. With roughly 2.1B Android devices worldwide at the time, “one in every 3 or 4 Android phones had its product.” One function, driven all the way through.
• The person sitting next to him in a startup competition was 杨璐玉 of Musical.ly, the predecessor to TikTok—his English name was apparently “Louis.” Over lunch, they discussed: “I’ve reduced 3 or 4 buttons down to 3, and I’m thinking about how to turn them into 1.” Alipay’s home screen had 10 buttons at the time. 傅盛 saw the potential and invested, with a stake of either 10% or 20%; he no longer remembers the exact figure. The second lesson was “predict, break through with a single point, and go all in”: predict that the early market will expand, establish a foothold with one point, then deploy all your resources. “Given that I carry this methodology, it shouldn’t be surprising that I come up with new ways to play. It’s still the result of accumulated preparation finally paying off.”
• Cheetah went all in on AI around 2018. After spending several more years learning, by June 2022—before ChatGPT 3.5 had launched—he was already building AI products with reasonably good profitability, purely as a side business. That also explains why he does not raise money: “My revenue is already at a decent scale. Even if you gave me money, I wouldn’t know how to spend it.”

12. A Methodology of Luck: The Most Important Thing Is Showing Up

• He calls himself someone who has “ridden luck all his life,” but offers a formula for expanding the surface area of luck: “What’s the most important thing when you go out into the world? You have to go out. Once you show up, your luck can improve.” He went to Beijing alone for an internship in his junior year, seeking out teachers and department heads at his university one by one. Later he dared to enter 傅盛’s startup competition. “Even though I thought what I had built was bad at the time, if you dare to show up and expose yourself to more variables, eventually some of those variables will work in your favor.”

13. The Turpentine Moment: Buying Google on First-Hand Feel, Then Selling Everything

• The host invoked a Picasso anecdote: critics gather to discuss form and deeper meaning, while artists gather to discuss where to buy the cheapest turpentine. That is how he positioned 刘小排’s public account—as pure front-line tool intelligence. The strongest example came on May 6, when Gemini 2.5 Pro 0506 launched. “When it first launched, they hadn’t even changed the version number. It was actually 032-something. After testing it, I found that the model was simply the strongest in the world; other people just didn’t know it.” Google was trading at just over $130 that day. He bought heavily; a few months later it was above $200.
• The logic was straightforward: “Other investors don’t write code, so they have no way to judge what makes the model good. They might even conclude from media coverage that Google is overvalued.” After building his company, he mostly liquidated his stock holdings: “It also frees up my attention. Trading stocks can make money, but it doesn’t give me the same satisfaction as building products.” Koji echoed the view: 80% of his usage had shifted from GPT to Gemini 2.5 Pro. “At first I thought GPT’s accumulated memory meant I could never move house in this lifetime.”

14. Model Division of Labor: Claude Works, GPT Solves the Hard Problems, Gemini Acts as Architect

• He explains a phenomenon “many people can’t figure out”: Claude has little profile on coding leaderboards, but programmers love it. “What do people actually use to code every day? Isn’t it mostly simple tasks? More than 90% of coding work may involve interfaces, logins, and payments. That’s exactly where Claude Code is strongest. Competition leaderboards, by contrast, reward getting hard problems right.”
• In competitive algorithms, research, and mathematics—areas where the training data barely exists—“you still have to say GPT is the strongest,” including o3, o4-mini, and GPT-5. When working on more difficult algorithmic code, he switches back to GPT. Gemini’s role is architect: “It has had 1T tokens of context for a long time, while Claude only has 200K. Put much of the project into its context, discuss the overall architecture, produce a technical plan, then hand it to Claude Code to execute.”

15. The Operating Principle: Save Your Time, Not Its Tokens

• He admits that from Claude Code’s perspective, his usage is inefficient. That is precisely the point: “I don’t care how much time or how many tokens it uses. I only care about the time it saves me.” For example, to translate an entire site into 32 languages, the token-efficient approach would be to ask it to write a translation script first. “But that’s too much trouble for me. I just ask Claude to translate everything. I don’t care how it does it. I care about translation quality. I’m going to sleep.”
• The old era was defined by ACM competitions, Dijkstra, and dynamic programming, where the goal was execution efficiency. “That’s the mindset the previous era left us with. Today it has changed—my own time is the most valuable thing. There’s no need to drag myself into the details of algorithmic execution efficiency.”
• His top technique for avoiding a codebase “pile of shit”—a year ago everyone asked about Cursor, now they ask about Claude Code—is to “write a serious requirements document.” He writes 1,000-2,000 Chinese characters in Feishu, with prototype images, then pastes it in and says: “Don’t write code yet. First ask me whether I’ve explained the requirements clearly. Explain your understanding back to me in your own words. Then make a technical plan, and ask me about anything that requires a decision.” He discusses it for 5-6 rounds until “nothing has been left unconsidered,” then starts. It is like the joke about asking a husband to “buy a watermelon, and if you see a bun shop, buy 2,” only to have him return with 2 watermelons: “That’s because your wording was ambiguous.”
• The 3 features he recommends most are background commands, for 7×24 background tasks; subagents, where “you create a separate agent and train it to do one type of task very well,” mainly for non-coding work; and output style’s learning mode. “It won’t finish the task for you. It leaves part of it for you to fill in, which is perfect for beginners.”

16. Citely: A 20-Year Business Built Around Academic Citations

• The incubated product Citely, at citely.ai, was disclosed publicly for the first time. Its founder is a professor at the Communication University of China. “I dare to disclose this because other people can’t build it.” Every pain point centers on citations: students using DeepSeek to write papers and fabricating references, a problem that hurts both teachers and students; automatically adding authoritative citations to a handwritten paper; and checking whether a passage is supported by actual research. “You ask GPT, but you don’t trust what it tells you.”
• His conclusion is worth noting: “This can be a 20-year business.” Citation has been an industry for 200 years. “As long as I keep focusing on real pain points around academic papers and repeatedly solve them with the latest technology, it isn’t a wrapper product. I can do this until retirement.” The founder is himself an academic, so “from day one you already know where the users are. Just go ask them for money.”

17. The Case for Profits Below the Top Tier—and the Final Principle: You Need a Mine to Roam Free

• He disagrees with prominent investors who say AI products do not make money. “It’s the top-tier AI products that don’t make much money, especially those that took investment. They have to do PR; their users were created through PR. The six startups that raised billions don’t make money. The products that make money are people like us, in the middle or below the middle. There are no marketing costs, labor costs are low, and yes, I must be profitable from day one. There are still infinite niches in AI.”
• The only difference from the previous era is that individuals have become stronger. “You already have a very capable programmer in your hands. Their coding is definitely better than 99% of humans.” The strategy is to keep finding pain points and validating them with MVPs: “If it works, make it bigger. If it doesn’t, abandon it. It’s like playing a game: if you lose this round, play another.” He cited a recent post by Sam Altman: “Making SARS products with AI today is a lot like fast fashion.”
• His goal in entrepreneurship is “narrative self.” “When I’m 80 or 90 and looking at my grandchildren, am I going to tell them, ‘Your grandfather owns 5 houses in Beijing’? Of course not. I’ll say, ‘Did you hear about this thing your grandfather built?’” He has not achieved it yet, but that is his dream.
• He closes with his only advice to people who want to act: “It is not for you, it is you.” “It’s not about what this tool can do for you. It’s about what you want to do, then deciding which parts you can outsource to it. It is part of you.” First, combine AI with a business or product that can support you. “Life is the wilderness. You need a mine before you can roam free”—once you have that mine, you can do what you want and explore the possibilities of this era with greater ease.