After Opus 4.5, AI Is Giving Rise to "Giant-Style Hundred-Person Companies"

In the six months since Opus 4.5's release, the Agent track has undergone a quiet reshuffling. The conversational differences between models have become nearly imperceptible to average users—what's truly widening the gap between models is the Agent Loop layer. Claude Code was the first to make a model's long-horizon task planning and tool-calling proactivity visibly apparent; Opus 4.5 pushed language models into the Agent era; and the "wrapper" controversy has gradually been disproven by product strength.

Alongside this main narrative runs a more hidden thread: when the production subject shifts from "a person assisted by tools" to "an Agent capable of independent work," the products, organizations, and commercial defaults built around "humans as production subjects" all need to be re-examined—the meaning of open source, how ToB pricing works, whose voice products should follow, the relationship between organizational scale and output value. Every proposition that seemed settled in Chapter 1 (ChatGPT 3.5) has been reopened in Chapter 2 (Opus 4.5).

Cherry Studio is a noteworthy case study along this hidden thread. Around it, we spoke with founder Yinsen and 5Y Capital investor Steven about some judgments rarely seen in the industry today.

Guests

Yinsen | Founder, Cherry Studio

Steven | Investor, 5Y Capital

Host

Qianqian | Head of Brand, 5Y Capital

Highlights

• The only assets that truly constitute moats for AI companies today are great teams and great users.
• Open source in the AI era is engineering infrastructure for asymmetric iteration—the most cutting-edge startups in the US are already using work trials to replace interviews; Claude Code and OpenClaw can ship three versions per week because two-thirds of their roadmap comes from open-source community issues.
• The product methodology of "finding the greatest common denominator" has failed in the Agent era—Doubao's product managers shouldn't listen to the median user's demands in a normal distribution, but to those "capable of writing issues."
• Opus 4.5 opened the most important paradigm of language model Chapter 2: the flywheel of strongest model × strongest use case × strongest user is now fully spinning.
• "Never having been burned by Chinese ToB" is actually an advantage for the new generation of AI entrepreneurs—Sell Work is replacing SaaS.
• The most formidable companies of the next era won't be one-person companies, but "giant-scale hundred-person companies."

Below are our edited core arguments:

01 The Adoption Layer: Independent Value Obscured by the "Wrapper" Label

Qianqian: How should we understand "wrapper"—is it an undervalued user need?

Yinsen: I think it's severely undervalued. We now divide it into three layers—the first is the model layer, the second is today's Agent runtime layer, and above that we don't call it the "application layer" but the Adoption layer—it encompasses both "application" and "how it's used," two critical elements.

Today everyone suspects whether the application layer is too thin, whether it'll be swallowed by model OEMs. But what's overlooked is the "how it's used" layer. The more capable models become, the more Agents can do, the more challenges they encounter in a complex world—like an intern whose tasks are all simple; but once they become an independent business lead, challenges multiply.

Model OEMs focus on universality, the greatest common denominator; those making "wrappers" focus on real practices in specific scenarios—embedding Agents into the daily workflows of an engineer, a legal counsel, an IT colleague. The actual landing of these usage scenarios is the value wrappers bring.

Steven: My perspective is somewhat different. My observation is—the best wrappers have the best users. Among AI companies today, aside from model companies, the only assets that truly qualify as moats are two kinds: exceptionally good teams, and exceptionally good users. Everything else is hard to accumulate sustainably.

Whether Monica previously, or Cherry Studio today, they both corresponded to the most advanced, best AI users at their respective moments.

02 Open Source's New Meaning: From Ecological Contract to Asymmetric Iteration

Qianqian: How do you view open source? Does it mean the same thing in the AI era as before?

Yinsen: Open source has different meanings in different eras. Before AI, open source could form excellent ecosystems, avoiding reinventing the wheel—Linux systems, various foundational component libraries all grew this way. But after AI, the value of code openness itself is declining—vast amounts of code are now AI-generated, and AI is trained on massive amounts of open-source code. But another dimension of value is rising—the value of building in public. Open code means smoother feedback channels between you and the market, and users.

Steven: I've actually invested in numerous open-source projects, from Camel—the first Multi-Agent framework six months after ChatGPT's 2023 release—to Dify, Moonshot AI, and Cherry Studio. The biggest lesson I've learned from positive examples like DeepSeek and Anthropic is that the team is the core asset of an AI company. Not ARR, not users, not network effects, not funding—the core asset is the team.

And open source is the best means to recruit people suited to each team. It validates several crucial assumptions—people attracted through open source have good culture fit with existing teams, and very high agency. This is what AI-native teams value most today.

A more cutting-edge form is that the most frontier startups in the US are already experimenting with replacing job interviews with work trials. Having candidates work together for three to five days, or even one day, is more meaningful than traditional interviews. I want to know how well you use Claude Code, not how well you talk.

Open source has another layer of meaning that's only emerged in the past six months—asymmetric advantages in product iteration. The release velocity of Claude Code and OpenClaw teams is astonishing, capable of three versions per week, all solving fairly important features.

Why? Because their roadmaps are built on open-source community issues. Assuming I have 100 issues, if I'm Boris (Claude Code's engineering lead), I only need to abstract the 10 most priority-worthy problems and have Claude Code solve them. You don't need to define "what I should do for users" like a traditional product manager—two-thirds of things derive from the open-source community.

I have a hot take—today Doubao's product managers probably shouldn't listen to 3-sigma Doubao user demands in a normal distribution; this doesn't help with building Agents.

03 Why You Shouldn't Listen to "Median User" Feedback

Qianqian: How large is the user value variance you've observed?

Yinsen: In the Agent era, user value variance is extraordinarily large. We have users consuming $1,000 in tokens monthly; we also have users who can't finish $10 in two months. Their value is completely asymmetrical.

Previously there were only two kinds of software—software nobody uses, and software full of problems. Because once software is used intensively, it inevitably exposes numerous issues. But much software never reaches this point. Open source has advantages here.

More crucially—after future coding capabilities are partially replaced by AI, it becomes less scarce. What's truly scarce is demand—high-quality demand originating from real users. Whoever has demand can iterate their product at high speed; products without demand, AI also doesn't know how to write, and fall into death spirals.

Steven: This is vastly different from previous internet products. Today, among users, the value between user and user has become very uneven. Internet products have always sought the greatest common denominator, finding commonalities, solving them, capturing maximum traffic. But in the Agent era, this variance appears extraordinarily large.

This was also what attracted me to Cherry Studio initially. When we first met, when founders came to me for pitch meetings and screenshared from their laptops, I'd see an icon I didn't recognize—and discovered it was Cherry Studio. A wave of China's best AI users were almost all on Cherry Studio.

Yinsen: Cherry Studio doesn't do coding scenarios—that scenario's know-how is very concrete, and definitely model OEMs' business, with no opportunity for third-party products. We focus on non-coding general tasks, so we early on gathered a group of high-productivity-value knowledge workers without technical backgrounds.

Initially hardcore technical-background users adopted first, thoroughly used the product, then wrote tutorials on Xiaohongshu, WeChat official accounts, and Douyin—these platforms most easily reach knowledge workers. Later a category of users emerged—they had large text workloads that other products couldn't satisfy, and Cherry Studio's knowledge base combined with DeepSeek solved their problems. Once a user reported an issue, I asked him to screenshot, and he said he didn't know how. But he was urgently using Cherry Studio. In that moment I realized—the user profile was shifting.

04 How the Claude Code Flywheel Spins Up

Qianqian: Is "Claude Code is hard to surpass" an industry consensus?

Steven: I don't think so. If we agree "it's hard to surpass Claude Code," then various Agent Infra projects need not be done—just wait for Claude to teach you. This certainly isn't consensus, but this is my somewhat realistic, conservative view.

Doing better than Claude Code is hard, not because Boris has better intelligence than us—it's because he has the strongest model. The strongest model generates the strongest use cases, and these use cases correspond to the strongest users—his flywheel is fully spinning. People catching up with various agent framework designs actually find it hard to catch up; this is a very realistic advantage.

Yinsen: Using browser engines as analogy—from Mozilla, to Safari's WebKit, to Chrome's Blink—today no browser doesn't use the Blink kernel. This doesn't mean Google's technical capability is strongest, but that it's become the de facto standard. All web frontends are designed and developed targeting Blink's best rendering effect. It underwent massive use case validation and became the world's standard.

Steven: This timing reminds me of Douyin and TikTok in 2017–2018. Back then people tried various ways to make better short-video interactions, to innovate, but after ten years we discovered—the interaction paradigm Musical.ly invented was simply the best. Musical.ly's flywheel truly spun up when its interaction paradigm met ByteDance's algorithm. Before that, because algorithms differed, this flywheel didn't spin so perfectly. Today's turning point is products like Claude Code after Opus 4.5's release—this loop is fully spinning.

ChatGPT 3.5 was Chapter 1, Opus 4.5 is Chapter 2—by Chapter 2 it's thoroughly the Agent era. But if we continue using Sonnet 3.5—this model can't run. No matter how good your harness, how brilliant your context management, it's meaningless.

Yinsen: The right users will naturally appear on the right model and right framework. Just as with Douyin's swipe interaction and recommendation algorithm, users who become addicted to it will inevitably emerge.

05 Chinese ToB's Non-Consensus Window: Not Having Been Burned Is an Advantage

Qianqian: How do you view Cherry Studio's chosen ToB track?

Steven: Cherry Studio chose ToB, and Chinese ToB—put these two words together and they'd scare off the vast majority of China's investment market. I was ignorant and fearless; I'd never lost money on Chinese ToB before. So I still have strong interest in this proposition. I believe AI's variable for ToB efficiency tools is larger than the previous cloud computing era, and there may even be considerable business model variables.

Previously SaaS wasn't, in outcome terms, such a suitable business model for China. But AI's use-case-oriented, more obvious value proposition—I think there may be completely unexpected surprises in China.

And Yinsen's previous experience also satisfies many of my fantasies—one hypothesis of mine is: this person preferably hasn't done much Chinese ToB before, like me; it's best if neither of us has been burned.

Yinsen: I indeed have little ToB experience; my only somewhat ToB stint was at Yitu, because previous-generation AI companies were mainly ToB/ToG. I have no obsession with doing ToB, nor do I reject it. My logic is simple—the greatest transformation this wave of AI brings is productivity transformation. Enterprises are the most direct, most important carriers of social productivity; productivity-enhancing technology should be applied within organizations as production subjects.

Steven: From business path design, the Sell Work approach is more suitable for startups—it has asymmetric advantages relative to giants. Giants' original business ceilings far exceed VC imagination. Take ByteDance: if not for AI, last year it might have had tens of billions in profit. If a startup monopolized China's entire legal services market, how meaningful is that to ByteDance? But for a startup, the incremental change this brings is enormous. Healthcare, accounting, various professional services—the same logic applies.

For giant ToC platform companies, this new incremental is a rather鸡肋 existence—but for startups, it's alpha. Asymmetry precisely hides in this "not worth their while." You must believe that technology itself brings more dividends than market environment. We should first satisfy technology dividends, embrace technology dividends, then satisfy user needs.

Qianqian: How to understand "unleashing technology dividends"?

Yinsen: In March–April 2025, one point made me anxious—I found that differences in conversational experience between different models were shrinking, increasingly unperceived. Whether using the strongest GPT 4.5, Claude 3.7, or domestic Qwen, the experience difference was already very small. Meaning Cherry Studio's product form at that time could no longer bring out models' longest advantages. Benchmarks were failing—the product wasn't maximizing model dividend release.

Our true solution came after seeing Claude Code's release. Once Claude Code emerged, everyone was exposed—the differences in models' long-horizon task planning, tool-calling proactivity, and final delivery quality were vast. Then we saw the real opportunity lay in Agents—Agents could once again widen the performance gap between models, once again release model dividends.

06 Organizational Entropy Reduction

The Hundred-Person Giant Company Is Coming

Qianqian: What's the ultimate mission of ToB tools for an organization?

Yinsen: From a biological perspective, from single cells to multicellular organisms, to tissues, to organs, to motile individuals—system complexity rises higher and higher, as does the ceiling. But massive information transmission and material exchange within become significant overhead—in some sense less efficient than single-cell individuals, yet an inevitable choice.

Organizations are the same. From two people to 200 to 2,000, the time and cost each person spends on alignment grows rapidly. Organizations divide into countless layers, essentially just passing information up and down—collecting execution results to report upward, receiving upper-level goals to decompose downward—a process夹杂大量信息失真和个人主观色彩, making organizational entropy increase very fast.

But at the AI stage, passing information up and down can naturally be solved by AI. Past OKR decomposition to execution layers fundamentally didn't match, became unrecognizable, and nobody had capacity to check every individual. But Agent AI can do this—it can align very seriously, very clearly. In the future Agents will directly exchange information through context, with very high bandwidth and very little distortion, greatly reducing entropy increase trends.

Using the Industrial Revolution as analogy—the handicraft era relied on manual polishing, with very poor product consistency. But in the assembly line era, the production subject wasn't people but machines, with tolerances controlled within very narrow ranges. In the future knowledge work will also shift from handicraft to industrialization, and the entropy increase within will be reduced to unprecedented levels.

Qianqian: Will future super-large organizations be mainstream?

Yinsen: I think there will still be companies with such high output value, but they won't need so many people; they'll become very flat. Perhaps a formerly 100,000-person company can manage with 1,000. Smaller companies can produce higher output value.

Entropy reduction occurs not only within organizations, but between organizations—such as supply chains. A simple structural component, behind it has mold factories, injection factories, material factories, painting factories—previously this was a networked large organization with even lower communication efficiency than within organizations. If the entropy increase of such collaborative networks becomes very small, it means social wealth development speed will become very fast.

Steven: Because of Matthew effects in this world, the most formidable companies of the next era definitely won't be one-person companies—but there will be very formidable hundred-person companies, giant-scale hundred-person companies.

Yinsen: Individual companies have numerous limitations—organizations lack resilience, are fragile, hard to grow into towering trees. But hundred-person companies, several-hundred-person companies will definitely produce very formidable giants. I see two aspects. First is positive attitude toward AI—look at actions, not what they say. The quantitative metric is TOKEN consumption. Many say bosses who use token consumption as KPI are xx—I am one. But you must first get token consumption up, first don't pursue which portion is effective, which ineffective—everyone will surely摸索出 different手感 through the process. Like writing calligraphy, first don't talk technique, first finish writing three large vats of water.

More quantitatively, look at how much TOKEN budget and quota companies allocate to employees. Many companies say they embrace AI, love AI, but won't spend ¥100 to buy Claude for employees—such companies aren't truly embracing AI.

Second is looking at what kinds of companies will increase their share of social production value after AI brings changes to organizational form, organizational layers, organizational scale. What we currently see is that medium and small-to-medium companies are growing fastest. For giants, essentially 100,000 people isn't their shortcoming; their shortcoming is business ceiling. But a 1,000-person company's business ceiling may be very high, only lacking people—this portion's increment is very fast.

5Y Capital seeks, supports, and inspires lonely entrepreneurs, providing support from spirit to all operational management. We believe that if the you others see as crazy begins to be believed in, the world will become refreshingly different.

BEIJING·SHANGHAI·SHENZHEN·HONGKONG

WWW.5YCAP.COM