AI: Money Printing Machine or Money Shredder? Pricing Crawfish by Token—you might not be able to afford to raise them!

Recently, "raising lobsters" has become a topic that's hard to avoid.

Some people are sharing their "lobster diaries" on their social circles, nearly a thousand people are queuing downstairs at the Tencent tower, and some are willing to spend 500 yuan to have someone come and install it, just to get it running as soon as possible.

Many people are tempted, driven by the same stimulus: the imagination of "one person equals a team". Especially with the story of Fu Sheng during the Spring Festival holidays, which was regarded as the emotional trigger of this wave: bedridden from a fracture, he exchanged 1,157 messages totaling 220,000 words with OpenClaw in 14 days, turning a "novice who couldn't even check his contacts" into an automated team made up of 8 Agents—even to the point that the "lobster" automatically posted tweets at 3 a.m. and achieved a million views.

It sounds like a machine that can print money automatically, but in the most sensational narratives, what's most easily ignored is the bill, threshold, and risk.

You think you're installing the same "lobster," but you might not be; you think the cost is in the software, but the real expense is in the token; you think the 500 yuan you spent is for "early use, early profit", but in three months it might depreciate to zero.

Fu Sheng's "myth" does exist, but it's very expensive

In a recent livestream, Fu Sheng admitted that he spends over $100 a day on OpenClaw—that's about 700 yuan.

This is just for personal use. He has connected the lobster to the company's Feishu, cultivated multiple Agents such as consultants, copywriters, and operations officers, and the entire system's daily token consumption is much higher.

More importantly, the results are not "install and you get it." Fu Sheng's output is based on financial investment, top-tier model configuration, and dozens of days of intensive training; his core advantage is the ability to precisely break down tasks, write Skill documents, and design feedback loops—the real "secret weapon" is human capability, not the default "lobster" out of the box.

Using such a case as an ordinary person's template is like persuading someone who just got their driver's license to take an F1 car onto the track—the outcome is rarely "take off," but more often "crash."

Token Black Hole: The little lobster you can't afford

What many didn't anticipate is that OpenClaw's cost isn't in the software itself, but in the backend model calls.

It's inherently a "token black hole"—every task consumes a large amount of tokens in interactions with the backend large model, and once task chains lengthen, tool calls increase, or memory is enabled, costs rise rapidly.

An ordinary chatbot conversation just uses a few hundred tokens; OpenClaw performing the same task may require millions of tokens.

Users report that searching for information and writing a 2,000-word document can burn through 7 million tokens; running a simple crawler test actually uses 29 million tokens; cases burning through 50 million tokens in a day are common.

One SaaS company even gave every employee a "lobster subsidy," with ordinary employees consuming 150 yuan in tokens daily, and the tech team up to 1,000 yuan.

More hidden is OpenClaw's "heartbeat mechanism"—even with no real output, the system still automatically consumes about 145 yuan in calls per day, which translates to a monthly loss of over 5,000 yuan.

This "money-burning" experience strongly resonates with overseas users as well.

There has been users whose token costs reached $1,500 in a week, about 10,381 yuan.

X user @Kekius Sage posted with a meme: "This is my first day using OpenClaw, I'm really feeling old." He just had the lobster read a few recent research papers, and his daily bill shot up to $22.1. He sighed in the post that if this goes on, next week it might cost $1,000 a day.

 The "lobster" you installed may not be the same one at all

Many don't know that OpenClaw's capability ceiling depends on which deployment path you use.

OpenClaw has accumulated nearly 250,000 stars on GitHub, which is exciting. It's not just a chatbot, but a digital executor with "hands": it can open browsers, read and write files, send emails, and manage social accounts. That's all true.

But in reality: your deployment method determines its actual abilities. The mainstream paths currently on the market are roughly four:

Local dedicated hardware (typically standalone Mac Mini): highest capability ceiling, can read files, control browsers, manage calendar and emails, with more complete context. But the price is steep: a hardware investment starts at 4,000 yuan; to run local large models and avoid API dependence, high-end workstations easily break 100,000 yuan. Even with cloud models, the ongoing cost remains API call fees.

Cloud server deployment (Aliyun, Tencent Cloud, etc one-click solutions): with monthly fees from tens to a hundred yuan, looks affordable, but the "lobster" is locked in an "empty room"—no files, no authorized accounts, work is naturally cut in half, more like an upgraded chatbot.

Direct installation on a personal computer: lowest threshold, highest risk. Since it shares permissions with the OS, any misconfiguration or malicious Skill injection affects not just the "lobster," but your data and accounts. Security agencies say over 40,000 OpenClaw instances are exposed to public networks, more than 60% contain exploitable vulnerabilities.

Vendor-hosted solutions (like KimiClaw, MaxClaw): plug-and-play, deployment barrier almost zero; but "visibility" and autonomy are limited by platform rules, with capability ceilings and data autonomy boxed in.

Big companies' pricing strategies make the illusion of "low threshold" exceptionally convincing:

KimiClaw's Vivace membership is 199 yuan monthly; Aliyun packs top models into subscription, Pro first month 39.9 yuan, renewal 100 yuan, and provides 90,000 API requests per month; MiniMax's MaxClaw puts OpenClaw directly on the cloud, claiming no need to prep your own server, basic features available via "MiniMax Agent Basic Edition Subscriber".

But if you do the math, these packages are essentially "limited buckets." 90,000 API requests seem plenty, but once the lobster starts multi-Agent collaboration or long task chains, it runs out in days. Exceed the quota and pay per use, prices jump back to market rates, and end-of-month bills are often shocking.

Is this lobster qualified as a proper AI assistant?

OpenClaw went from a weekend experiment in November 2025 to today, not even four months. It's a fast-iterating but still rough open-source project, clearly short of being a mature product.

Main known defects include: simple tasks processed overly complex at times; unexplained interruptions during task execution; unstable memory function, sometimes "forgetting" previous preferences; efficiency between token consumption and output is far from optimized. Among thousands of Skills on ClawHub, hundreds contain malicious code.

Binance founder Zhao Changpeng (CZ) posted a brief but powerful message on X: "Claims that after installing the lobster, you don't need to do anything. Afterwards, all your time goes to tweaking the lobster that can do nothing." This nails the real user experience for most—"freeing up your hands" often turns into "occupying your hands".

Security risks are the sword of Damocles over every user's head. Security research shows more than 40,000 OpenClaw instances exposed on public networks, 63% with exploitable vulnerabilities. Even Google, Anthropic, Meta have banned OpenClaw internally—not because of technical issues, but because current security can't keep up with capability expansion.

The most alarming case comes from Meta's superintelligent team AI alignment lead Summer Yue. She gave the lobster a seemingly simple command: "Check my inbox and suggest which emails can be archived or deleted." The lobster started bulk deleting, safety limits failed, only a physical shutdown stopped it.

Even Musk is worried about OpenClaw's safety: "People are giving their whole life's root privileges to OpenClaw."

OpenClaw founder Peter Steinberger has himself been honest: "If you don't understand command line, this project is too risky for you." This sentence is worth every potential lobster installer thinking twice about.

Don't let AI anxiety become the sickle that reaps you

 Back to the fundamental question: Why are nearly a thousand people willing to queue downstairs at Tencent Tower, and some willing to pay 500 yuan for home installation?

Someone put it simply: "It's all about information gaps, targeting those who fear 'doing-it-yourself' and 'code.'" Basically, many aren't buying the lobster, but buying certainty to counter AI panic—they're afraid they're missing out on what was the Internet 20 years ago, or bitcoin 10 years ago.

This anxiety is rational to a degree. OpenClaw does prove the exciting possibility of "AI as more than a chat window, but a true doer." But possibility isn't reality, prototype isn't product, and others' success stories aren't your use case.

Yet, anxiety-driven consumption has always been the favorite prey of "shovel sellers"—from 500 yuan home install fees, to one-click deployment packages, to Skill packs, API price differences, enterprise AI training courses... the whole industry chain has clearly built itself around your insecurity.

The lobster won't automatically become a good employee, just as a good computer won't make you a great programmer. AI is an amplifier, and it amplifies human ability first. Fu Sheng turned a lobster into a team because he already had the base ability to build a team, not just because he wanted to "install it first and see".

Maintaining independent judgment amid the noise, refusing to let "fear of lagging behind" push you to make irrational choices—that's what every irreplaceable human individual should prioritize as a core asset in this AI wave.

Risk warning and disclaimerThe market has risks and investment requires caution. This article does not constitute personal investment advice, nor does it consider individual users' special investment goals, financial conditions, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their specific situation. Investments based on this are at your own risk.