Deconstructing Anthropic: The Best AI Company, Possibly Also a Type 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 Anthropic's most capable generation of models to date. 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 intriguing than the model itself 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 scales, 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 the competition in the AI industry may never have been just about models. Often, 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 computing power supply keeps up, it is likely to reach 100B by the end of the year and 200-300B next year, directly matching Meta's scale. In the secondary market, its valuation has already reached 1 trillion USD, surpassing OpenAI.

We spent considerable 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 in 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 person at Anthropic 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 different; it was the only one among the big three to abandon multimodal early on, and it has never talked about architectural innovation or 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.

Regarding why coding is so important, the market is now clear on three core points:

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 of results and a short feedback loop, allowing user data to better feed back into model training.

3. Coding is the core accelerator for AGI development. Leading AI labs have already entered this acceleration loop, with this year's model improvements occurring faster than in the past year.

The final results confirm 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, it turns out to be half foresight and half luck.

Anthropic's early financing was quite challenging. 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. 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, explaining 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 changed; financing became smoother, and the company had more resources, so the coding line was no longer mentioned, and they went on to create a more general model base.

The turning point occurred after the explosive popularity of ChatGPT. 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 B-end. This strategic shift was still cautious and empirical, not a reckless gamble.

When training Claude 3, Anthropic began to consciously strengthen its coding capabilities and received good market feedback on Sonnet 3.5. Subsequently, it increased investment while seeking validation, gradually solidifying its judgment on coding potential, 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 from star researchers have repeatedly claimed that scaling laws have hit a wall and that the marginal returns of pretraining have peaked. From our communications with various researchers, Anthropic has consistently been the most confident in scaling laws among all labs and has solidly executed pretraining and data without dispersing energy on new paradigms.

Looking back, this was indeed the right approach. The leap in Claude's capabilities largely came from solid investments in pretraining.

The Founder's Personality

But this raises another question: why does Anthropic consistently make decisive trade-offs in several key directions and maintain its composure?

First, there is the limitation of resources; Anthropic's historical financing amount is only about one-third 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. Dario himself was the core research lead for GPT-3 and 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 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 overturning overnight, I began to focus on my own bets. I don't know if we are right, but to be honest, even if we are only right 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 close to Sam, we see that:

• 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 seeds in multiple places and betting in parallel," which led OpenAI to grow countless side projects.

• Sam does not come from a technical background, so his judgments on technical directions are not as strong as Anthropic's, relying more on the team to push from the bottom up. Sam leverages his strengths in resource management to provide ammunition to each team.

• His VC background makes Sam particularly fond of breakthrough fancy ideas. Thus, OpenAI's culture 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 and are abandoned after launch.

• Both Sam and Mark Chen (Chief Research Officer) have personalities that only say yes and do not say no. For side projects, as long as the team works hard, resources will still be provided from above.

When OpenAI's forces are 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 gives us an insight that 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 essence of all the entrepreneurial experiences distilled from over 400 biographies he has read, he replied: Focus.

Great entrepreneurs are often not well-rounded top students but rather extreme obsessives. They identify 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 push them to the extreme, even to the point of absurdity for competitors.

It should be clarified that many people think they are focused, but they do not truly understand the meaning and cost of focus. The so-called focus essentially breaks 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, while 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, users did not need a larger portal but rather the ability 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. Its business model was similarly focused; while Yahoo had dozens of monetization methods, Google concentrated all its efforts on "search keyword bidding," spending 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 job-hopping 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%, which is 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 when OpenAI was at its peak, and Anthropic had not yet emerged. If we look at recent news over the past two years, Anthropic's talent attraction and stability become even more apparent. For example, a recent popular post on Twitter 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 greatest 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 it, because when people talk about culture or values, it often feels vague, assumed to be just a slogan. However, when you overlay all firsthand information and public interviews, it can be quite striking.

Three Characteristics of Anthropic

If we break it down specifically, three characteristics that set Anthropic apart 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-image: 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 realities.

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 about this 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 most companies' logic, 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 in line with this principle, such as their governance structure designed with a non-profit trust in mind, research on explainability, various investments in safety, including the recent willingness to sacrifice a $200 million contract with the U.S. Department of Defense due to value conflicts, etc., which I will not elaborate on here.

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 remarkable thing is that Frontier AI is a place where star culture and resource struggles can easily emerge. AI researchers are among the smartest and highest ego individuals in the world; their natural pursuit is to propose different solutions, establish their own factions, and achieve fame, but resources are very limited, so departmental conflicts often occur.

Daniel Freeman, who jumped from Google to Anthropic, said that other model companies internally feel like individual fiefdoms, each managing their own affairs and secretly competing, but he "has never felt that way at Anthropic."

Rahul Patil, former CTO of Stripe, mentioned after joining Anthropic last fall that what struck him the most was the culture here. It is hard to imagine that such a group of intelligent 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 do so, without ego.

3. A Strong Humanistic Background

A writer from The New Yorker spent a few months deeply following Anthropic 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 typical Silicon Valley elites or traditional impressions of technical geeks; they have a bit of bookishness, a bit of nerdiness, and a bit of idealism. Many people give the impression of having grown up in families of writers and poets.

This can be seen to some extent in the naming of Claude models: Haiku, Sonnet, Opus, corresponding to the concise haiku, Shakespeare's 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 somewhat illustrates the differences.

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 referenced 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 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 rate 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 handle 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 for cultural interviews, asking 15-20 scenario questions in an hour. Based on the interview questions circulating 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 ensure 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 admit one's ignorance and mistakes.

• Can you handle complexity? Many issues dealt with 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 decisions and how one 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 top 10x developers. Rahul Patil mentioned that before joining Anthropic, he had a long conversation with the then CTO of Anthropic. The latter not only did not persuade him to join 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 come.

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

In contrast, OpenAI, after growing larger, no longer conducts dedicated cultural interviews, which has reportedly led to some management issues. This was particularly evident during Meta's recent hiring spree. Faced with Meta's exorbitant packages, OpenAI's response was more like a market norm: counter offers, retention bonuses, and canceling new employees' vesting cliffs to expedite stock vesting. Anthropic's response 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 final 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. A Culture of Context Sharing

Anthropic has very high information transparency internally.

First, Dario himself actively, frequently, and repeatedly provides meaning. 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 had some epiphany). He stands in front of the entire company for an hour, usually accompanied by a three to four-page document, discussing everything from company direction and product strategy to industry changes, and then directly answering 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 not calculatedly but says exactly what he thinks."

In addition to all-hands meetings, he frequently writes many things in his Slack channel, unadorned, 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 what should be prioritized. Thus, in a complex and changing situation, each individual can make relatively consistent distributed decisions.

Moreover, this transparency is not a one-way transmission but can be challenged. Someone might listen to Dario's sharing in 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 that involves everyone. Many people at Anthropic have their own notebook channels, somewhat like personal Twitter feeds, recording 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 treasure trove where many things unfold.

Thus, Anthropic seems to have cultivated a very good alignment soil within the company, where everyone's projects, viewpoints, and ideas 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 do not often eat together. The result is that some researchers from other companies have expressed regret that all key know-how is scattered 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 goes against commercial common sense: it has seven founders, and Dario resolutely decided to give everyone equal shares rather than taking a larger portion for himself.

At the time, everyone advised him that this would be a disaster; otherwise, the leadership 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 any one 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 design for governance rights 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 less likely to dilute its original culture.

In contrast, OpenAI's executive team has been very unstable; 11 founding team members have left in succession, leaving only Sam Altman, Greg Brockman, and Wojciech Zaremba still present. The newly appointed executive team is even less stable: since the beginning of 2026, the head of product has taken leave, the head of marketing has left for health reasons, the head of communications has exited, the head of operations has been reassigned, and the head of finance has also been marginalized...

4. Extremely Emphasizing One Team, Avoiding Factionalism

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

The most important work happens at the front lines. Because the frontline people 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 are proposed by frontline personnel rather than driven by executive roadmaps.

However, this also poses a problem; when judgment is decentralized, each team can easily cling to its own problem awareness and value function, growing into individual factions that pull against each other.

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 the founder's perspective, with everyone participating in a massive trade-off processing in their respective roles.

Thus, they emphasize one team and also 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," "architect vs. implementer."

This is 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, lacking much voice. When conflicts arise, research is also unwilling to cooperate with product.

In product innovation, OpenAI has a strong characteristic of being researcher-driven: often, a research team produces a new result, and the product team only receives an email and starts looking for a nail with a hammer. 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 actually 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 it can be viewed from two aspects:

1. The Requirements of the Business Itself

I remember two years ago listening to a talk by an HR leader from a top company, 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 directions of coding and agency, it may seem like a competition of model capabilities on the surface, but at a deeper level, it is actually a competition of engineering capabilities. It is not a problem that can be solved by a few geniuses having a moment of inspiration; rather, it involves a lot of dirty, fragmented, and detailed systems engineering.

The most critical barrier is data. Previous chat data was simply text data, but coding and agentic data are more complex; they not only include 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 to get right, but it does not create personal highlights like publishing a paper or launching a new product.

From feedback we received from 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 handle dirty work. OpenAI hires the top talents from the hierarchy of disdain, with good backgrounds and high aspirations; everyone naturally wants to make their own bets, and few are willing to deal with the mess or fill in data.

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 handling data personally every day, with extremely meticulous data cleaning that 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 of a research problem, but in everyday agentic tasks, they often do not perform as well as Anthropic, as the latter is more of an engineering problem, testing data, systems, and execution details.)

2. The Founding Team's Background

Company values can be said to be 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 greater than that of the former. We can take a simple look at Dario's experience:

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

Dario later joined OpenAI, where he was deeply involved in advancing 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, differences in organizational philosophy began to emerge between him and others at OpenAI.

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 commercial disagreement but a fundamental issue of values.

Greg and Dario's paths diverged over the years, with Sam Altman caught in the middle trying to mediate. Sam played one of his strongest abilities at this time, which was to make different factions feel that he was actually on their side. In the short term, this was a balancing act; in the long term, it was a depletion of trust. 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, with some people calling this small group "the pandas" because of his fondness for pandas. Their disagreements with OpenAI's leadership on issues like strategic choices and organizational governance grew larger, eventually developing into serious political struggles.

There was even a serious confrontation between the upper management. Sam accused Dario and Daniela (Dario's sister and one of Anthropic's later co-founders) of organizing negative feedback against him behind his back; the two denied it and called in the source of Sam's information for confrontation. The result was that the person claimed to be completely unaware of the matter, and Sam then turned around to deny having made the accusation.

This incident caused Dario and his sister to completely lose trust, and they ended up arguing.

There are many similar internal dramas; in short, Dario elevated the conflicts between the two sides to a moral crisis of trust. He felt that a company wielding such powerful technology must have leaders who are sincere and trustworthy. If the person at the helm is dishonest, it is helping to build 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 only due to Dario's inherent nature but also significantly shaped by his personal experiences of political struggles at OpenAI. He clearly understands how easily a group of high-ego smart individuals can split due to resource competition and value disagreements, so they instinctively built Anthropic in the opposite direction:

• Having seen how balancing acts can deplete 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 disintegration caused by 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 behind by OpenAI.

03 Conclusion

In summary, 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 seeking the next breakthrough super platform.

To clarify further, we can place several core dimensions of both companies side by side:

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

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

• Demand is exploding 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 continues to explore and bet 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.