Deconstructing Anthropic: The Best AI Company, Possibly Also a Kind of Organizational Invention

Source: Overseas Unicorn Celia

Today, Anthropic released its next-generation flagship model Claude Fable 5, as well as Claude Mythos 5, which is open to specific institutions.

From the data disclosed by the official sources, this is the most powerful generation of models that Anthropic has produced so far. Especially in software development, complex knowledge work, and long task processing, it has further widened the gap with the previous generation of products.

However, more than the model itself, what is more intriguing is why Anthropic has consistently been able to keep pace with industry changes over the past few years.

While most companies were still discussing parameter scale, it focused on coding; when the industry was competing for C-end traffic, it turned to the enterprise market; as more and more companies began to expand everywhere, it concentrated resources on a few key directions.

Looking back, Anthropic's rise does not seem like a technological miracle, but rather the result of a company adhering to a certain judgment over the long term. In the past year, Anthropic has grown from a company considered a follower of OpenAI to one of the most watched players in the AI industry. Revenue, valuation, and talent attraction are all rapidly climbing. Many attribute all of this to Claude.

But if we extend the timeline a bit, we might find that Claude is perhaps just the result. What is truly worth studying is why Anthropic has consistently been able to see important directions earlier than others, and why it has been able to maintain restraint in the face of all temptations.

The release of Fable 5 serves as a new footnote. It reminds people once again that competition in the AI industry may never have been just about models. Many times, what determines victory or defeat are strategy, organization, and what a company is willing to give up for what it values.

In the past year, Anthropic may be the most worthy company to study in the entire AI industry.

At the beginning of this year, it created the fastest explosive growth in human business history: ARR grew from 9B to 45B, and if the supply of computing power keeps up, it is highly likely that ARR will reach 100B by the end of the year and 200-300B next year, directly aligning with Meta's scale. In the secondary market, its valuation has already touched 1 trillion USD, surpassing OpenAI.

We spent a lot of time studying how Anthropic managed to rise from behind. Ultimately, to understand this company, the core is to grasp two points: one is strategic judgment, and the other is organizational culture.

Everyone should already have many fragmented understandings of this, but there is no complete picture, so this article attempts to provide a more detailed sorting and restoration. It hopes to explain some curious questions from the outside world from the perspectives of strategy and organization, such as:

• Why did Anthropic realize as early as 2021 that coding might be the most important direction?

• How did the personality differences between Dario and Sam shape the completely different strategic paths of the two companies?

• Why is Anthropic's talent turnover rate so low?

• Why does almost every Anthropic employee praise its culture? How is this culture maintained during the company's rapid expansion?

01 The Importance of Focus is Underestimated

First of all, strategically speaking, OpenAI has always seemed more like a company that wants everything.

In terms of model capabilities, OpenAI is making efforts in math, science, coding, reasoning, multimodal, architectural innovation, etc. In terms of products, Codex, browsers, robots, enterprise platforms, smart hardware, chips, and data centers are all being advanced simultaneously. It is said that the number of projects within OpenAI once reached about 300.

In contrast, Anthropic is completely the opposite; it was the only one among the big three to abandon multimodal early on and has never talked about architectural innovation, nor emphasized concepts like reasoning models, RL, or continual learning. It focuses solely on scaling language models, focusing only on coding as a direction, first mastering the most critical capabilities.

The market is now clear on why coding is so important, with three core reasons:

1. Coding is the pathway to everything. The vast majority of tasks in the digital world can be expressed through code.

2. Coding is the most suitable capability for model learning. It has strong verifiability, a short feedback loop, and user data can greatly feed back into model training.

3. Coding is the core accelerator for AGI development. Leading AI labs have already entered this acceleration loop, and this year, the progress of models in a single quarter is faster than in the past year.

The final result confirms that coding is indeed the most important direction, overshadowing everything else. OpenAI only woke up to this in March, cutting off side projects like Sora and elevating coding to the company's top priority.

How Did Anthropic Accurately Choose Coding?

We have always been curious: How did Anthropic manage to accurately choose coding from the start? Tracing back, we find that it is half foresight and half luck.

Anthropic's early financing was quite difficult. With limited funds, it had to move towards AGI in a more efficient way. It needed to tell a story about a vertical scenario to prove it could form a commercial closed loop. So they seriously studied that if they could only choose one direction, coding might be the best choice: first train a better coding model → provide it for customer use → obtain customer usage data in real engineering environments → feed back into model training. This could potentially form a flywheel.

Anthropic's growth leader once mentioned that he had seen an internal document written by the company's co-founders, which explained why they should focus on coding. The key point is that this document was dated 2021, long before anyone knew what the actual market opportunity for this direction was.

However, the situation later was that financing became smoother, and the company had more resources, so the coding line was no longer mentioned; they still went on to create a more general model base.

The turning point occurred after ChatGPT exploded in popularity. Anthropic realized that the C-end had already been seized by OpenAI, so it regretfully (but in hindsight, very fortunately) shifted the battlefield and turned its focus to the B-end. This strategic shift was still cautious and empirical, not a reckless gamble.

When training Claude 3, Anthropic began to consciously strengthen coding capabilities and received good market feedback on Sonnet 3.5. After that, it was a matter of increasing investment while seeking validation, and internally, the judgment on coding's potential gradually solidified, both in terms of commercial value and accelerating research. Thus, the team began to focus on this path, completely abandoning the C-end and even not diverting energy to multimodal.

In addition to market direction focus, it is also worth mentioning the steadfastness in technical routes. Over the past two years, external voices have repeatedly claimed that scaling laws have hit a wall, and the marginal returns of pretraining have peaked. From our exchanges with various researchers, Anthropic has consistently been the most confident in scaling laws among all labs and has done the most solid work in pretraining and data, without dispersing energy on new paradigms.

Looking back, this was indeed the right approach. A significant part of Claude's capability leap came from solid investment in pretraining.

The Founder's Personality

But this raises another curiosity: Why does Anthropic always manage to make decisive trade-offs in several key directions and maintain its composure?

First, naturally, there is the limitation of resources; Anthropic's historical financing amount is only about 1/3 of OpenAI's. But looking deeper, the strategic differences between the two companies are also closely related to the personalities and backgrounds of their founders.

Anthropic has four co-founders who were core authors of the scaling laws paper, and Dario himself was the core research lead for GPT-3. Before that, he had already spent ten years in the AI field, having firsthand experience of technological advancements in AI, making him more daring in judgment. Additionally, Dario is someone who does not experience FOMO at all; he has even been described as somewhat narcissistic and stubborn, rarely being swayed by market consensus.

In 2024, when Anthropic had not yet achieved explosive growth, he said something that I believe is very important for understanding this company:

"The deepest lesson I've learned in the past decade is that there will always be a so-called consensus in the market, but after seeing several instances of consensus flipping overnight, I began to focus on my own bets. I don't know if we are definitely right, but honestly, even if we are right only 50% of the time, that is already very valuable, as you are providing something that others do not have."

This is very different from Sam Altman. From our conversations with some people close to Sam, we see:

• Sam is recognized as one of the most ambitious founders in Silicon Valley, wanting everything from the start. Coupled with his past investment experience at YC, he is very familiar with the method of "planting multiple seeds and betting in parallel," which is why OpenAI has grown numerous side projects.

• Sam does not come from a technical background, so his judgments on technical directions are not as strong as Anthropic's, which is why he relies more on the team to push from the bottom up. Sam plays to his strengths in resource mobilization, providing ammunition to each team.

• His VC background makes Sam particularly fond of breakthrough fancy ideas. Thus, OpenAI's culture highly values 0 to 1 paradigm innovation but does not equally emphasize the continuous refinement from 1 to 10. Many product lines, such as Sora, Atlas browser, and Voice Mode, lack continuity; once launched, they are left unattended.

• Both Sam and Mark Chen (Chief Research Officer) have personalities that only say yes and do not say no. For side tasks, as long as the team works hard, the upper management will still provide resources.

When OpenAI's manpower is continuously diluted by various side projects, Anthropic can gain an advantage on the most critical battlefield through strategic competition.

The Brilliance of Strategy Lies in "Subtlety"

Anthropic's strategic focus has inspired us; the importance of focus is underestimated.

I recall a podcast I listened to last year, featuring David Senra, the host of the Founders podcast. For the past eight years, he has done almost one thing: studying a great entrepreneur each week. When asked what would be the one takeaway from all the entrepreneurial experiences distilled from over 400 biographies he has read, he answered: Focus.

Great entrepreneurs are often not well-rounded top students but rather extreme perfectionists. They can identify the one or two variables that are most important to them, such as Costco's pricing, Apple's design experience, or ByteDance's recommendation algorithm & data flywheel, and they will push them to the extreme at all costs, even to the point of absurdity for competitors.

It is important to clarify that many people think they are focused, but they do not truly understand the meaning and cost of focus. The so-called focus essentially needs to be broken down into two levels:

1. Judgment: Knowing what is most critical and daring to sacrifice everything else.

2. Pressure: Being able to invest overwhelming resources to penetrate the key elements.

The former is a cognitive issue, and the latter is a will issue; both are indispensable.

For example, when Google was founded, the consensus in the entire internet industry was that the future belonged to "portals." Giants like Yahoo were filling their homepages with more and more features—news, weather, shopping, games, horoscopes… every feature was seen as a lever to "increase advertising value." But Google believed that as information increased, what users needed was not a larger portal but to find the most relevant answers immediately.

Thus, while others wanted users to stay longer, Google wanted users to leave faster. At that time, Google's homepage was exceptionally clean, with nothing but a search box. In terms of business models, Yahoo had dozens of monetization methods, while Google focused all its energy on "search keyword bidding" for nearly a decade before seriously considering a second business line.

To this day, one of Google's tenets is "It's best to do one thing really, really well."

The core of strategy is not to clarify what you want to choose but to clarify what you want to give up. I believe most people do not say no enough.

02 Culture is the Biggest Secret Sauce

The most special aspect of Anthropic may not be its strategy but its organizational culture. Over the past six months, in the fierce competition for AI talent, Anthropic's talent turnover rate has been far lower than that of other AI labs.

The following two charts summarize talent movement data from 2021 to 2023. The first chart shows the proportion of people switching between various AI labs, and we can see:

• For every 10.6 people moving from DeepMind to Anthropic, only 1 goes back to DeepMind.

• For every 8.2 people moving from OpenAI to Anthropic, only 1 goes back to OpenAI.

The second chart shows the proportion of employees who remain with the company two years after joining. Anthropic's talent retention rate is 80%, the highest among leading AI labs at the time, slightly higher than DeepMind's 78%. As a younger, rapidly changing company, Anthropic has managed to achieve a higher retention rate than the established DeepMind, which is not easy. In contrast, OpenAI has only 67%.

It is worth noting that this data was collected while OpenAI was at its peak, and before Anthropic had fully emerged. If we look at the news from the past two years, Anthropic's talent attraction and stability become even more apparent. For example, a recent popular Twitter post highlighted that several CTOs from star companies were willing to jump to Anthropic to become ordinary technical employees (i.e., MTS, member of technical staff).

The biggest reason for this is often attributed to Anthropic's organizational culture. If you look at the podcasts recorded by Anthropic members, almost everyone mentions Anthropic's culture, with some even viewing this cult-like culture as Anthropic's biggest secret sauce.

"I really believe that culture is Anthropic's secret weapon; it is our most defensible and irreplicable asset. This is not something that happens naturally; the leadership has invested a lot in this."

------ Amol Avasare, Anthropic Growth Leader

If you don't approach this issue with a specific awareness, you might not notice this point because when people talk about culture or values, it often feels abstract, assumed to be a slogan. However, when you layer all the firsthand information and public interviews together, it can be quite striking.

Three Characteristics of Anthropic

If we break it down specifically, three characteristics that make Anthropic different from other AI labs are:

1. Mission-oriented

Anthropic's mission is "to ensure that the world can safely navigate the transition to transformative AI," meaning everything is centered around safety.

Many companies claim to be mission-driven, but Anthropic's seriousness about this reaches a somewhat religious level. It is a frontier lab with a strong moral self-imagination: it genuinely believes that AGI can save the world and also genuinely believes that AGI could destroy the world, and it tries to lead everyone to walk the narrow tightrope between these two possibilities.

Boris Cherny, head of Claude Code, once said: "At Anthropic, if you randomly ask someone in the hallway 'Why are you here?', the answer will always be safety."

He and product manager Cat Wu both left Anthropic for Cursor last year but returned within two weeks because they found themselves deeply missing the cultural atmosphere within Anthropic. The feeling of everyone purely striving for a greater mission. Some who were skeptical before joining Anthropic found that "Wow, the atmosphere inside is even more serious than what is said outside."

There have even been early employees who said in all-hands meetings that if Anthropic ultimately achieves its mission but the company itself fails, that would still be a good outcome. This statement explains a lot about Anthropic.

In the logic of most enterprises, commercial success is always the top priority, and the mission is merely for decoration. But what is most special about Anthropic is that there is indeed a group of people internally who place the mission above the company's survival. If we examine what Anthropic actually does, it is also consistent with its actions, such as their governance structure designed with a non-profit trust in power, research on explainability, various investments in safety, including the recent willingness to sacrifice a $200 million order from the U.S. Department of Defense due to value conflicts, etc., which I will not elaborate on one by one.

2. High trust, low ego

When we communicate with other leading labs, we often hear about internal politics and factional issues. Only Anthropic does not have this. On the contrary, everyone is very united and willing to help others.

The most amazing thing is that Frontier AI is a place where star culture and resource struggles can easily emerge. AI researchers are arguably the smartest and most high-ego group of people in the world; their natural pursuit is to propose a different solution, establish their own faction, and make a name for themselves, but resources are very limited, so departmental conflicts often occur.

Daniel Freeman, who moved from Google to Anthropic, said that other model companies internally feel like individual fiefdoms where everyone is secretly competing, but he "has never felt that way at Anthropic."

Rahul Patil, former CTO of Stripe, who joined Anthropic last fall, also mentioned that what shocked him the most was the culture here. It is hard to imagine that such a smart group of people can also be so humble at the same time. He gave a standard: if the company tells you tomorrow that the best position for you is not to continue as an executive but to become an IC (individual contributor) because that would be your greatest contribution to the mission, would you be willing? He believes 100% of Anthropic's people would agree, with no ego.

3. A Strong Humanistic Background

A writer from The New Yorker once did an in-depth follow-up at Anthropic for a few months and left two interesting descriptions of the people there:

• Bookish misfits

• A disproportionate number of Anthropic employees seem to be the children of novelists or poets.

In other words, the people here do not resemble the typical Silicon Valley elite, nor do they fit the traditional image of tech-savvy engineers; instead, they have a bit of bookishness, nerdiness, and idealism. Many people give the impression of having grown up in families of writers and poets.

This can be seen to some extent from the naming of Claude models: Haiku, Sonnet, Opus, corresponding to the concise haiku, Shakespearean sonnet, and classical large works. In contrast, OpenAI's GPT-4 / 4o / o1 are named by engineering numbers, and Google's Gemini Ultra / Pro / Flash are named after classic product lines. This says something.

Boris, head of Claude Code, also shared an interesting detail in a podcast: during his first lunch at Anthropic, he casually mentioned a very obscure book by hard sci-fi author Greg Egan. How obscure was that book? He had never met anyone who had read it before. He casually mentioned a plot point from the book, and surprisingly, everyone at the table picked up on it.

This left him greatly shocked and made him feel he had truly come to the right place. Sci-fi-loving bookworms often possess a grand sense of humanistic concern and historical responsibility, as well as better reasoning abilities regarding the butterfly effect. This consensus based on reading interests reassured him that this might be the best place to push the boundaries of AI.

How Culture is Institutionalized

The next question is, how is this pure, almost cult-like culture maintained?

After all, Anthropic is no longer a small AI lab; it is a large company with 3,000 people, and it has managed to maintain its cultural density while expanding at the fastest pace in history.

In this regard, Dario directly stated that he probably spends 1/3 to 40% of his time ensuring that Anthropic's culture is good. Even with countless tasks to do in technology, products, financing, and political-business relations, he believes that his higher-leverage work is to make Anthropic a place where top talent wants to work due to its high cohesion.

In terms of specific practices, there are several points:

1. Special Recruitment Standards

Anthropic's hiring approach is different from many AI labs.

On one hand, in terms of talent preferences, unlike most companies competing for big names, Anthropic prefers to hire underdogs. Rather than external labels, they value direct evidence of ability, such as "Have you done independent research, written truly insightful blogs, or made substantial contributions to the open-source community?" etc.

On the other hand, Anthropic has very strict cultural screening. They have a dedicated round of Cultural interviews during the hiring process, asking 15-20 scenario questions in an hour. Based on the interview questions that have circulated online, they focus on three points:

• Do you really prioritize the safety mission? A typical screening question is: If Anthropic decides not to release a model because it cannot guarantee safety, would you be willing to accept your stock going to zero?

• Are you a nice, low-ego person? This includes kindness, empathy, people skills, and the ability to acknowledge one's ignorance and mistakes.

• Can you handle complexity? Many issues handled internally at Anthropic are very complex and variable; they value whether a person has systematic thinking and can deeply reason about the second-order effects of things, considering how a decision will impact other aspects.

They spend a lot of time on "reverse screening" in recruitment, which has led them to genuinely give up many of the top 10x developers. Rahul Patil mentioned that before joining Anthropic, he had a long conversation with the then Anthropic CTO. The latter not only did not persuade him to come over but also spent two to three weeks discussing why he should not join Anthropic, kindly advising him that unless he was truly aligned with the culture and mission, it would not be worth it to join.

Thus, Anthropic's recruitment logic has never been about bringing in as many of the strongest people as possible but rather about filtering out unsuitable individuals as early as possible. "We are very good at screening out those who come for money and fame."

In contrast, OpenAI, after growing larger, no longer conducts dedicated cultural interviews, which has reportedly caused some management issues. This was particularly evident during Meta's recent hiring spree. Faced with Meta's exorbitant package, OpenAI's response was more like market convention: counter offers, retention bonuses, and canceling vesting cliffs for new employees to expedite stock vesting. Anthropic's response, however, was very Anthropic. They told employees that they came here primarily for the mission, not to continuously raise their prices in external bidding. They would not offer you a salary ten times higher than equally qualified colleagues just because Mark Zuckerberg happened to notice you; that would be unfair, and if you want to leave, then leave.

The outcome of this situation is also telling. OpenAI reportedly lost dozens of people, while Anthropic only lost 2, and those two had already worked at Meta for 6 and 11 years.

2. Culture of Context Sharing

Anthropic has very high information transparency internally.

First, Dario himself actively, frequently, and repeatedly provides meaningful input. He often holds all-hands meetings to share with everyone in the company, with a frequency of up to once every two weeks, called Dario Vision Quest (even Dario himself jokes that the name's evangelistic quality is too obvious, sounding like he went to the mountains and came back enlightened). He stands in front of the entire company for an hour, usually accompanied by a three to four-page document, covering everything from company direction and product strategy to industry changes, and then directly answers questions on the spot.

Many internal employees say he speaks very directly and honestly, "Dario is the most straightforward person I've ever met; he speaks without calculation, just saying what he truly thinks."

In addition to all-hands meetings, he frequently writes a lot in his Slack channel, unadornedly recording his thoughts: what has happened in the company recently, what he is worried about, and how he views issues that concern everyone. This culture allows everyone in the company to know how decisions are made and which matters should be prioritized. Thus, in a complex and changing situation, each individual can make relatively consistent distributed decisions.

At the same time, this transparency is not a one-way transmission but can be challenged. Someone might listen to Dario's sharing at an all-hands meeting, feel disagreement, and directly go to Dario's notebook channel to publicly say, "I disagree with your judgment," and then engage in a debate on the spot. Publicly challenging leadership is encouraged.

Furthermore, this writing culture does not belong solely to Dario but is a thinking mechanism involving everyone. Many people at Anthropic have their own notebook channels, somewhat like personal Twitter feeds, where they record what they are thinking, doing, and what progress they have made at any time. Others can subscribe, observe, or join discussions. Many employees have commented that they really like the company's writing culture; Slack is a huge repository where many things unfold.

Thus, Anthropic seems to have cultivated a very good alignment soil within the company, where everyone's projects, viewpoints, and thoughts are sufficiently transparent and fluid, with some even lamenting that financial data is also transparent.

(However, in contrast, technical confidentiality is maintained very strictly; it is said that some teams are even deliberately isolated and cannot dine together. The result is that some researchers from other companies have expressed regret that all key know-how is dispersed in different people's minds, making it impossible to piece together a complete picture by just poaching a few individuals.)

3. Seven Founders with Equal Shares, the Founding Structure Itself is a Cultural Mechanism

Anthropic's founding structure has a design that is quite contrary to commercial common sense: it has seven founders, and Dario resolutely decided to give everyone equal shares instead of taking a larger portion for himself.

At that time, everyone advised him that this would be a disaster; otherwise, the power would be ambiguous, and incentives misaligned, making it easy for the company to fall apart due to internal strife. But Dario believed that the company should revolve around the mission, not around a single founder, and that equal shares were the most undeniable evidence of this philosophy.

The seven co-founders had already worked together for many years and had a high level of trust in each other. Equal shares are essentially not a governance design but a proof of commitment, a mechanism for cultural diffusion. The seven co-founders act as seven cultural replication nodes, each projecting values to a broader audience along different lines. This way, even as the company expands, it is not easy to dilute the original culture.

In contrast, OpenAI's executive team has been quite turbulent; 11 founding team members have left consecutively, and now only Sam Altman, Greg Brockman, and Wojciech Zaremba remain. The newly appointed executive team is even less stable: since the beginning of 2026, the product lead Fidji has taken leave, the marketing lead left for health reasons, the communications lead has exited, the operations lead has been reassigned, and the finance lead has also been marginalized…

4. Extremely Emphasizing One Team, Avoiding Factionalism

Anthropic's CTO once said in a podcast that AI labs, compared to traditional companies, are very bottom-up; it is a reverse pyramid organizational structure where power and creativity flow from the bottom up.

The most important work happens at the front line. Because those on the front line are closest to the emergent behaviors of AI. They run experiments daily and have the most intuitive understanding of what models can do. The vast majority of product ideas come from frontline personnel rather than being driven by executive roadmaps.

However, this also poses a problem; once judgment is decentralized, each team can easily cling to its own problem awareness and value function, growing into competing factions.

Anthropic's uniqueness lies in its early recognition that since judgment must be decentralized, it is even more important to actively foster unity. Dario does not want safety to only say that safety is the most important, while product only says that product is the most important, pushing all conflicts up to the higher-ups for resolution. One of his core management philosophies is to distribute trade-offs to each individual, allowing everyone to have a bit of a founder's perspective, with everyone participating in the same massive trade-off processing in their respective roles.

Thus, they emphasize one team and design various systems to weaken the boundaries between responsibilities, such as not distinguishing titles below the executive level, uniformly calling everyone members of technical staff, deliberately downplaying identity definitions like "researcher vs. engineer," "senior vs. junior," and "architect vs. implementer."

This stands in stark contrast to OpenAI, which has always had a stronger researcher culture, with a clear "hierarchy of disdain" existing internally: Researcher > Research Engineer > Software Engineer. As a result, products are often dominated by research, receiving little voice. When conflicts arise, Research is also unwilling to cooperate with Product.

In terms of product innovation, OpenAI has a strong characteristic of being researcher-driven: often, a new achievement from the research team is only communicated to the product team via email at the last minute, who then scramble to find a fit. In contrast, at Anthropic, the product and model teams are more closely integrated, allowing products to more effectively influence and define model capabilities.

This is also one reason why OpenAI's product strength is not as strong as Anthropic's.

Two Origins of Culture

The next question is, why has Anthropic formed this unique organizational culture? Perhaps we can look at it from two aspects:

1. The Requirements of the Business Itself

I remember listening to a talk by an HR leader from a top company two years ago, which left a deep impression on me and made me think deeply about what organizational culture really means for the first time. The essence of organizational culture is: the behavior patterns of employees are a key factor that helps the company achieve success.

Thus, the first principle of organizational culture is that the nature of the business determines the organizational culture.

In the AI competition, a core moat is enabling "smart people to do dirty work." Especially in the direction of coding and agentic tasks, it may seem like a competition of model capabilities on the surface, but at a deeper level, it is a competition of engineering capabilities. It is not the kind of problem that can be solved by a few geniuses having a flash of inspiration; rather, it involves a lot of dirty, fragmented, and detailed systems engineering.

The most critical barrier is data. Previous chat data was just simple text data, but coding and agentic data are more complex; they include not only dialogue records but also the tasks themselves, environment setups, execution trajectories, and the entire evaluation and verification system.

This involves a lot of dirty work, which is crucial, but it does not lend itself to becoming an individual's highlight moment like publishing a paper or launching a new product.

From our feedback gathered through discussions with some researchers, one of OpenAI's core issues today is that it struggles to organize hundreds of the strongest individuals to diligently work on data and do the dirty work. OpenAI hires the top talents from the hierarchy of disdain, with good backgrounds and high aspirations, and everyone naturally wants to make their own bets, while few are willing to take on the messy work of data cleanup.

OpenAI has been so successful in the past because it indeed gained a significant lead through some core paradigm breakthroughs, but as Yao Shunyu said in a recent interview: "The era of individual heroism is over," and "AI does not require much brainpower… the most important trait is reliability and attention to detail."

At this point, it becomes clear that Anthropic's low ego, strong cohesion, and mission-driven atmosphere amplify its advantages significantly. It is said that Anthropic's co-founder Jared Kaplan also leads his team in data handling daily, with data cleaning done extremely meticulously, something no other company can match.

(This also explains a phenomenon: OpenAI's models are the strongest in competitive coding challenges because these tasks are more research-oriented, but in everyday agentic tasks, they often lag behind Anthropic, as the latter's tasks are more engineering-oriented, testing data, systems, and execution details.)

2. The Founding Team's Background

Company values can be said to be a part of the founders' values. More accurately, the founders' values often come from two sources: one part is what the founders originally believe, and the other part is what they have deeply despised in the past.

The former determines what you want to become, while the latter determines what you absolutely do not want to become.

Anthropic clearly has both, and the shaping power of the latter may be even greater than that of the former. We can take a brief look at Dario's experience:

Dario first encountered AI in Baidu's AI lab, where he observed scaling laws for the first time and gradually became a firm believer in scaling laws.

Dario later joined OpenAI, where he was deeply involved in the advancement of the GPT series. OpenAI once allocated 50%-60% of the company's total computing power to him, allowing him to lead the GPT-3 project. However, because Dario is a person with distinct values and personal opinions, the differences in organizational philosophy between him and others at OpenAI began to gradually surface.

For example, Greg Brockman once proposed a shocking idea: in the future, AGI could be sold to nuclear powers in the UN Security Council. Dario nearly resigned on the spot; to him, this was not just a business disagreement but a fundamental issue of values.

For years, Greg and Dario's paths have been misaligned, with Sam Altman caught in the middle trying to mediate. Sam played to his strength in balancing, making both sides feel that he was actually on their side. In the short term, this is a balancing act; in the long term, it is a trust overdraft. Later, everyone realized that what Sam promised Dario and what he promised Greg were fundamentally different.

Gradually, Dario formed a tight-knit alliance within the company, and some people, because he liked pandas, referred to this small group as "the pandas." Their disagreements with OpenAI's leadership on strategic choices and organizational governance grew larger, eventually escalating into serious political struggles.

There was even a serious confrontation among the senior management. Sam accused Dario and Daniela (Dario's sister and a later co-founder of Anthropic) of organizing negative feedback against him behind his back; the two denied it and called in the source of Sam's claims for confrontation. The result was that the person had no idea about the matter, and Sam turned around to deny having made the accusation.

This incident caused Dario and his sister to completely lose trust, and they had a falling out on the spot.

There are many similar internal dramas; in short, Dario elevated the conflicts between the two sides to a moral crisis of trust. He believed that a company wielding such powerful technology must have leaders who are sincere and trustworthy. If the person at the helm is not honest, it is building a dangerous direction.

Thus, Dario ultimately left OpenAI with some core colleagues from the GPT-3 project and founded what is now Anthropic.

Therefore, the culture at Anthropic today is not just because Dario is inherently this way; more importantly, it is shaped by his personal experiences of political struggles at OpenAI, making him acutely aware of how easily a group of high-ego smart individuals can split due to resource competition and value differences. Consequently, they instinctively built Anthropic in the opposite direction:

• Having seen how balancing acts can erode trust, they emphasize authenticity and transparency more;

• Having witnessed intensified political struggles, they encourage everyone to address conflicts upfront and discuss them early;

• Having seen organizational collapse due to ideological differences, they set strict cultural screening;

• Having observed power struggles among superstars, they emphasize low ego and do not favor big names.

The organizational culture at Anthropic today is largely a reaction to the experiences left by OpenAI.

03 Conclusion

To summarize, Anthropic and OpenAI are actually two companies with quite different backgrounds. The former is an idealistic, mission-driven, and highly cohesive cult-like organization, while the latter is ambition-driven, multi-line expanding, and constantly searching for the next explosive point super platform.

To see it more clearly, we can juxtapose several core dimensions of the two companies:

However, although we have discussed many advantages of Anthropic, it is difficult to conclude that one culture necessarily outweighs the other, and it is also hard to predict the battlefield three months from now. The AI world changes too quickly, and OpenAI is currently being underestimated by the market, for example:

• Coding is already a known factor, and OpenAI is likely to catch up; a clear trend is that developers are migrating from Claude Code to Codex;

• Demand has exploded far beyond everyone's expectations, and computing power is becoming the new decisive factor, while OpenAI locked in computing resources far exceeding Anthropic's early on;

• OpenAI's culture of open exploration has its own significant advantages, and OpenAI is also continually exploring and betting on new paradigms more aggressively, with the next leap potentially turning the situation around.

It can only be said that looking back from 2026 at the past three years, Anthropic has indeed left a memorable example for the entire industry:

In the AI era, winning does not necessarily rely on greater ambition, more exploration, and stronger talent. Sometimes, winning can also come from the opposite: fewer bets, lower ego, and a naive mission.