Over the past year, Anthropic has likely been the most noteworthy company to study in the entire AI industry.
At the beginning of this year, it created the fastest explosive growth in human business history: Annual Recurring Revenue (ARR) grew from 9 billion to 45 billion. If computing power supply could keep up, ARR would likely reach 100 billion by year-end, and 200-300 billion next year, putting it on par with Meta's scale.
On the secondary market, its current valuation has already touched 1 trillion dollars, surpassing OpenAI.
We spent considerable time researching how Anthropic managed to catch up from behind.
Ultimately, to understand this company, the core lies in understanding two points:
One is strategic judgment, and the other is organizational culture.
Many should already have fragmented knowledge about these aspects, but not a complete picture. Therefore, this article attempts to provide a more detailed overview and reconstruction.
We hope to explain some of the questions outsiders are curious about 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 two companies' completely different strategic paths?
• Why is Anthropic's talent attrition rate so low?
• Why does almost everyone at Anthropic praise its culture? How is this culture maintained during the company's rapid expansion?
01. The Importance of Focus is Underestimated
First, strategically speaking, OpenAI has always been more like a company that wants everything.
In terms of model capabilities, math, science, coding, reasoning, multimodality, architectural innovation, etc., OpenAI is pushing forward on all fronts.
In terms of products, Codex, browsers, robotics, enterprise platforms, smart hardware, chips, and data centers are also being advanced simultaneously. It's said that the number of internal projects at OpenAI once reached about 300.
Anthropic is the complete opposite. They were the only one among the 'Big Three' to abandon multimodality early on, never talked about architectural innovation, never emphasized concepts like reasoning models, RL, continual learning, etc. They only focused on scaling language models and concentrated solely on the coding direction, aiming to break through the most crucial capability first.
Regarding why coding is so important, the market is now clear. The core reasons are threefold:
1. Coding is the path to everything. The vast majority of tasks in the digital world can be expressed through code.
2. Coding is the most suitable ability for model learning. The results are highly verifiable, the feedback loop is short, and user data can feed back into model training to a greater extent.
3. Coding is the core accelerator for AGI R&D. Top AI labs have already entered this acceleration cycle; the progress made by models in one quarter this year is faster than the progress made in an entire year before.
The final outcome confirmed that coding is indeed the most important direction, overshadowing all others.
OpenAI didn't wake up to this until March, cutting 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 pinpoint coding from the start?
Tracing back, we found it was half foresight, half luck.
Anthropic's early fundraising was once very difficult. With less money, they had to advance towards AGI in a more efficient way.
They needed to first tell a vertical story to prove they could form a business loop. So they seriously researched at the time: if they could only choose one direction, coding might be the best choice: train a better coding model first → provide it to customers → obtain customer usage data from real engineering environments → feed back into model training. This could potentially form a flywheel.
Anthropic's Head of Growth once mentioned seeing an internal document written by a company co-founder about why they should focus on the coding direction. The key point is that this document was dated 2021, far earlier than anyone knew the actual market opportunity of this direction.
But later, fundraising became smoother, the company gained more resources, and the coding thread was no longer mentioned. They still went to build a more general model foundation first.
The turning point happened after ChatGPT went viral. Anthropic realized the C-end was already preempted by OpenAI, so they quite regrettably (but, in hindsight, exceptionally luckily) shifted the battlefield, turning their focus to B2B.
This strategic pivot was overall cautious and empirical, not a resolute all-or-nothing gamble.
When training Claude 3, Anthropic consciously began strengthening coding capabilities and received good market feedback on Sonnet 3.5.
After that, it was a process of doubling down while seeking proof. Internally, they gradually solidified their judgment on coding's potential, both in commercial value and research acceleration. So the team started focusing on moving forward along this path. During this time, they not only completely abandoned the C-end but also didn't even divert energy to multimodality.
Besides focusing on the market direction, it's also worth mentioning their perseverance on the technical path.
Over the past two years, star researchers externally repeatedly said scaling laws had hit a wall, and the marginal returns of pretraining had peaked. From our exchanges with researchers from various labs, Anthropic has always been the lab that most believed in scaling laws and did the most solid work on pretraining and data, without getting distracted by new paradigms.
Looking back, this was also correct. A significant part of Claude's capability leap came from solid investment in pretraining.
The Founders' Personalities
But this raised another curiosity: Why has Anthropic always been able to make decisive trade-offs on several key directions and maintain perseverance?
First, naturally, resource constraints. Anthropic's historical fundraising is roughly only one-third of OpenAI's. But looking deeper, the strategic differences between the two companies are also closely related to the founders' personalities and backgrounds.
Four of Anthropic's co-founders were core authors of the scaling laws paper back then. Dario himself was the most core research lead for GPT-3 and had already been in the AI field for a decade before that, having a firsthand sense of AI technological progress, making him more willing to make judgments.
Furthermore, Dario is someone who is completely not fomo (fear of missing out), even described as somewhat narcissistic and stubborn, rarely led by market consensus.
In 2024, when Anthropic was far from achieving explosive growth, he said something that I still think is very important for understanding this company, roughly meaning:
The deepest lesson I've learned over the past ten years is that there will always be a so-called consensus in the market. But after seeing the consensus flip overnight several times, I started focusing 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, it's already very valuable because you provide something others don't.
This is very different from Sam Altman. From our exchanges with some people close to Sam:
1. Sam is one of the most ambitious founders recognized in Silicon Valley, wanting everything from the start. Plus, his past experience investing at YC made him very familiar with the method of 'scattering seeds and making parallel bets,' so OpenAI grew countless side projects.
2. Sam doesn't come from a technical background, so his judgment on technical directions is not as strong as Anthropic's, relying more on the team to push forward bottom-up. Sam plays to his strength of resource acquisition, delivering ammunition to each team.
3. His VC background makes Sam particularly fond of breakthrough, fancy ideas. So OpenAI's culture highly values paradigm innovation from 0 to 1 but doesn't equally emphasize continuous polishing from 1 to 10. Many product lines like Sora, the Atlas browser, Voice Mode, etc., lack continuity; they are launched and then left unattended.
4. Sam and Mark Chen's (Chief Research Officer) personalities are such that they only say yes, not no. For side tasks, as long as the team pushes hard, the leadership will still provide resources.
While OpenAI's manpower was continuously diluted by various side projects, Anthropic could gain an advantage on the most critical battlefield through 'horse racing tactics.'
The Brilliance of Strategy Lies in 'What You Omit'
Anthropic's strategic focus gave us an insight: the importance of focus is underestimated.
I recall a podcast episode I listened to last year, where the guest was David Senra, host of the 'Founders' podcast. For the past 8 years, he has done almost one thing: studying one great entrepreneur each week.
When asked: if he compressed all the entrepreneurial lessons distilled from the over 400 founder biographies he read into just one thing, what would it be?
He answered: Focus.
Great entrepreneurs are often not all-around top students but extreme fanatics. They identify the one or two most important variables for themselves, like Costco's price, Apple's design experience, ByteDance's recommendation algorithm & data flywheel, and then push them to the extreme at all costs, even to a degree that seems absurd to competitors.
It's important to clarify here: many people think they are focused, but they don't truly understand the meaning and cost of focus.
So-called focus essentially needs to be broken down into two levels:
First, judgment – knowing what is most crucial and daring to sacrifice everything else.
Second, pressure – being able to invest overwhelming resources to break through the key element.
The former is a cognitive problem, the latter a willpower problem; both are indispensable.
For example, when Google was founded, the consensus across the internet industry was – the future belonged to 'portals.' Search giants like Yahoo were making their homepages increasingly cluttered with news, weather, shopping, games, horoscopes... each feature seen as a lever to 'increase advertising value.'
But Google believed information would keep increasing. What users needed wasn't a bigger portal, but to find the most relevant answer immediately.
So, while others wanted users to stay longer, Google wanted users to leave faster. At the time, Google's homepage was exceptionally clean, featuring nothing but a search box.
The same went for the business model. Yahoo had dozens of monetization methods. Google put all its energy into the single mechanism of 'search keyword bidding' and did it for nearly ten years before seriously starting a second business line.
To this day, one of Google's ten core philosophies is 'It's best to do one thing really, really well.'
The core of strategy is not figuring out what you want to choose, but what you are willing to give up. I think most people don't say 'no' enough.
02. Culture is the Biggest Secret Sauce
The most special thing about Anthropic might not be its strategy, but its organizational culture.
Over the past six months, in the fierce AI talent war, Anthropic's talent attrition rate has been far lower than other AI labs.
The following two charts summarize talent mobility data from 2021-2023.
The first chart shows the proportion of people moving between various AI labs. We can see:
• For every 10.6 people going from DeepMind to Anthropic, only 1 goes the opposite way.
• For every 8.2 people going from OpenAI to Anthropic, only 1 goes the opposite way.
The second chart shows the proportion of employees still at the company two years after joining.
Anthropic's talent retention rate is 80%, the highest among top AI labs at the time, slightly higher than DeepMind's 78%.
As a younger, fast-changing company, Anthropic managed to achieve higher retention than the veteran DeepMind, which is not easy.
In contrast, OpenAI's is only 67%.
It's worth noting that this data was compiled when OpenAI was at its peak and Anthropic had not yet shown its potential.
Looking at news from recent years, Anthropic's talent attraction and stability are even more evident.
For example, a recent popular Twitter post: CTOs from several star companies willingly jumped to Anthropic to become ordinary technical staff (i.e., MTS, member of technical staff):
The biggest reason for this is often attributed to Anthropic's organizational culture.
If you listen to podcasts by Anthropic members, almost everyone mentions Anthropic's culture. Some even see this cult-like culture as Anthropic's biggest secret sauce.
"I really think culture is Anthropic's secret weapon, our most defensible, un-copyable thing by others. This isn't natural; leadership has invested a lot in this."
——Amol Avasare, Head of Growth at Anthropic
If you don't look with this specific question in mind, you might not notice this point, because when people talk about culture or values, it often feels abstract, assumed to be just a slogan. But if you overlay all the firsthand information and public interviews, it's quite striking.
Three Traits of Anthropic
To break it down specifically, three traits that make Anthropic very 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 safety comes first.
Many companies claim to be mission-driven, but Anthropic's seriousness about this borders on religious.
This is a frontier lab with a strong moral self-perception: it genuinely believes AGI can save the world and also genuinely believes AGI could destroy the world, and it's trying to lead everyone across the narrow tightrope between the two.
Boris Cherny, Head of Claude Code, once said: "At Anthropic, if you stop anyone in the hallway and ask 'Why are you here?', the answer will be 'safety.'"
Both he and Product Manager Cat Wu left Anthropic last year to join Cursor but returned after less than two weeks because they found themselves deeply missing Anthropic's internal cultural atmosphere—that feeling of everyone purely striving for a greater mission.
Some were skeptical before joining Anthropic but found that, "Fuck, the atmosphere inside is even more serious than what's said outside."
There were even early employees saying at all-hands meetings—if Anthropic ultimately achieves its mission but the company itself fails, that's still a good outcome.
This sentence explains many things about Anthropic.
In the logic of most enterprises, commercial success always comes first; mission is just for decoration. But the most special thing about Anthropic is that there is indeed a group of people internally who place the mission ahead of the company's survival.
Examining what Anthropic actually does shows they walk the talk: their governance structure design with non-profit trustees in control, research on interpretability, various investments in safety, including recently sacrificing a $200 million U.S. Department of Defense order due to value conflicts, etc. We won't elaborate on all these here.
2. High trust, low ego
When we communicate with other frontier labs, we often hear about internal politics and factionalism. Only Anthropic doesn't have this. On the contrary, people are very united and willing to pave the way for others.
The most magical thing here is that Frontier AI is a field where star culture and resource struggles easily emerge. AI researchers are among the smartest, highest-ego people in the world. They naturally seek to propose different solutions, establish their own factions, and gain fame, but resources are very limited, so departmental conflicts always happen.
Daniel Freeman, who moved from Google to Anthropic, said other model companies internally feel like separate, secretly competing fiefdoms, but he has "never felt this at Anthropic."
Rahul Patil, former CTO of Stripe, who joined Anthropic last autumn, also mentioned being most impressed by the culture here. It's hard to imagine such smart people could simultaneously be so humble.
He gave a criterion: If the company told you tomorrow that the position most suitable for you isn't continuing as an executive, but becoming an IC (Individual Contributor), because that's your biggest contribution to the mission, would you be willing? He believed 100% of people at Anthropic would do it, with no ego.
3. A strong humanistic undertone
A New Yorker author did a months-long deep follow-up inside 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.
That is to say, the people here don't quite resemble typical Silicon Valley elites or traditional technical/engineering types. Instead, they are somewhat bookish, a bit nerdy, a bit idealistic. Many give the feeling of having grown up in families of writers and poets.
This is somewhat evident from the Claude model names: Haiku, Sonnet, Opus, corresponding to concise haiku, Shakespearean sonnets, and classical lengthy works.
For comparison, OpenAI's GPT-4 / 4o / o1 are named with engineering codes; Google's Gemini Ultra / Pro / Flash are classic product line names. This says something.
Boris Cherny, Head of Claude Code, also shared an interesting detail in a podcast:
At his first lunch at Anthropic, he casually mentioned a very obscure book by hard sci-fi author Greg Egan.
How niche was this book? He had never met anyone who had read it before.
He casually mentioned a reference from the book at the table, and everyone at the table actually got it.
This shocked him greatly and made him feel he was in the right place.
Sci-fi-loving nerds often have a certain grand humanistic concern and sense of historical responsibility, and also possess better reasoning abilities regarding butterfly effects.
This consensus based on reading tastes made him more confident 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's a large company of 3,000 people, and it's maintaining its cultural density while expanding at the fastest pace in history.
On this, Dario directly said he probably spends about one-third to 40% of his time ensuring Anthropic's culture is good.
Even though there are countless things to do technically, product-wise, fundraising-wise, and in government/business relations, he believes his higher-leverage work is making Anthropic a highly cohesive place where top talent loves to work.
In terms of concrete practices, there are a few points:
1. Special Hiring Criteria
Anthropic's hiring approach differs from many AI labs.
On one hand, in talent preference, unlike most companies competing for big names, Anthropic prefers hiring underdogs. More than external labels, they value direct evidence of ability, e.g., "Have you done independent research, written truly insightful blogs, made substantial contributions to the open-source community," etc.
On the other hand, Anthropic has very strict cultural screening. They have a dedicated Cultural interview round, asking 15-20 scenario-based questions in an hour.
Based on interview questions circulating online, they focus on three points:
(1) Whether you truly prioritize the safety mission.
The most typical screening question is: If Anthropic ultimately decides not to release a model because safety cannot be guaranteed, would you accept your stock options becoming worthless?
(2) Whether you are a nice, low-ego person.
Including kindness, empathy, people skills, ability to admit ignorance and mistakes.
(3) Whether you can handle complexity.
Many problems handled internally at Anthropic are very complex and variable. They highly value whether a person has systems thinking, can deeply reason about second-order effects, and think about how a decision affects other aspects.
They spend a lot of time on "reverse screening" in hiring and have indeed given up many top-tier 10x developers for this. Stripe's former CTO Rahul Patil mentioned that before joining Anthropic, he had long discussions with Anthropic's then-CTO.
The other party not only didn't persuade him to come but spent two to three weeks repeatedly discussing why he shouldn't join Anthropic, kindly discouraging him unless he was truly aligned on culture and mission, otherwise it wouldn't be worth it.
So Anthropic's hiring logic has never been to recruit as many strong people as possible, but to filter out unsuitable people as early as possible. "We are very good at filtering out people who come for money and fame."
In contrast, OpenAI stopped doing dedicated cultural interviews after the company grew, reportedly causing some management issues.
This was evident during Meta's talent-poaching round last year. Facing Meta's sky-high packages, OpenAI's reaction was more like market convention: counter offers, retention bonuses, removing vesting cliffs for new hires to accelerate stock vesting.
Anthropic's reaction was very Anthropic. They told employees: You're here first for the mission, not to keep raising your price in external bidding wars.
We won't offer you ten times the salary of equally excellent colleagues just because Mark Zuckerberg happened to pick you; that's unfair. If you want to leave, then leave.
The outcome of this incident is also telling. OpenAI reportedly lost dozens of people, while Anthropic only lost two, both of whom were former Meta employees with 6 and 11 years of tenure.
2. Context Sharing Culture
Anthropic has very high information transparency internally.
First, Dario himself proactively, frequently, and repeatedly provides meaning. He often holds all-hands meetings to share with everyone, as frequent as every two weeks, called Dario Vision Quest (even Dario himself jokes this name sounds too preachy, like he went to the mountains, inhaled something, and had an epiphany).
He stands in front of the whole company speaking for about an hour, usually with a three- to four-page document, covering everything from company direction, product strategy, to industry changes, then takes questions on the spot.
Many internal employees say he speaks very directly and candidly. "Dario is the most straightforward person I've met. His words aren't calculated; he genuinely says what he thinks."
Besides all-hands meetings, he frequently writes many things in his Slack channel during regular times, completely unpolished, recording his musings: what's happening lately, what he's worried about, and his views on issues everyone cares about.
Such a culture lets everyone in the company know how decisions are made and what should be prioritized. Thus, in a complex, changing situation, each individual can make relatively consistent distributed decisions.
At the same time, this transparency is not one-way indoctrination; it can be challenged. After hearing Dario's sharing at an All Hands, someone who disagreed directly went to Dario's notebook channel publicly saying "I disagree with your judgment" and then engaged in a debate on the spot. Publicly challenging leadership is encouraged.
Furthermore, this writing culture doesn't only belong to Dario; it's a thinking mechanism involving everyone.
Many at Anthropic have their own notebook channels, somewhat like personal Twitter feeds, recording what they are thinking, doing, and any progress at any time. Others can subscribe, observe, or join discussions.
Many employees have commented they like the company's writing culture; Slack is a huge treasure trove where many things unfold.
So Anthropic seems to have cultivated a good layer of alignment soil internally. Everyone's projects, views, and ideas are sufficiently transparent and fluid. Someone once even remarked that financial data is transparent.
(However, conversely, confidentiality on the technical side is very tight. It's said some groups are even deliberately isolated and can't easily have meals together.
The result is that researchers from other companies lament that all key know-how is scattered in different people's minds; it's impossible to piece together a full picture just by poaching a few people.)
3. Seven Founders with Equal Equity; Founding Structure Itself is a Cultural Mechanism
Anthropic's founding structure has a design that goes against common business sense: it has seven founders, and Dario insisted at the time on giving everyone equal equity, not taking more for himself.
At the time, everyone advised him this would be a disaster, leading to ambiguous leadership, misaligned incentives, and the company easily falling apart due to infighting.
But Dario believed the company revolves around the mission, not one founder, and equal equity is the most unforgeable evidence of this belief.
They had already worked together for years, highly trusting each other. Equal equity essentially wasn't a governance design but proof of commitment, a cultural diffusion mechanism.
Seven co-founders are like seven cultural replication nodes, projecting values to broader groups across different lines. This way, even as the company expands, it's less likely to dilute the initial culture.
In comparison, OpenAI's executive layer has been very turbulent. Eleven founding team members left one after another; now only Sam Altman, Greg Brockman, and Wojciech Zaremba remain.
The new executive layer is even more unstable: from the beginning of 2026 to now, Product Head Fidji took leave, Marketing Head left due to health reasons, Communications Head was ousted, Operations Head was reassigned, Finance Head was sidelined...
4. Strong Emphasis on 'One Team,' Avoiding Fiefdoms
Anthropic's CTO once said in a podcast that AI labs, compared to traditional companies, are very bottom-up; it's an inverted pyramid organization where power and creativity flow from bottom to top.
The most important work happens at the front lines because front-line people are closest to AI's emergent behavior. They run experiments daily and have the most intuitive understanding of what models can do. The vast majority of product ideas are pushed by front-line people, not driven by executive roadmaps.
But this also has a problem: when judgment is decentralized, each team easily guards its own problem awareness and value function, growing into factions pulling against each other.
Anthropic's specialness lies in realizing early: since judgment must be distributed, one must proactively create unity. Dario didn't want the safety team only saying safety is most important, product only saying product is most important, then pushing all conflicts up for leadership to decide.
A core management philosophy of his is to distribute trade-offs to each individual, letting everyone have a bit of a founder's perspective. Everyone is just participating in the same massive trade-off processing from their respective positions.
So they strongly emphasize 'one team' and use various institutional designs to weaken boundaries between roles. For example, below the executive level, there's no title distinction; everyone is uniformly called a Member of Technical Staff, deliberately weakening identity definitions like 'researcher vs. engineer,' 'senior vs. junior,' 'architect vs. implementer.'
This contrasts sharply with OpenAI, which has always had a stronger researcher culture, with a clear internal 'hierarchy': Researcher > Research Engineer > Software Engineer.
So product is often overshadowed by research, not gaining much voice. When conflicts arise, research is unwilling to cooperate with product.
In product innovation, OpenAI has a strong feature of being researcher-driven: often, the research team produces a new result, then the product team gets an email and starts looking for nails to hit with the hammer.
At Anthropic, product and model teams mesh more closely; product can more effectively influence and define model capabilities.
This is actually one reason why OpenAI's product strength is inferior to Anthropic's.
Two Origins of the Culture
The next question is: Why did Anthropic develop this unique organizational culture?
Perhaps we can look at it from two aspects:
I. Requirements of the Business Itself
I remember a share two years ago by an HR head of a leading large company, which left a deep impression and made me think deeply about what organizational culture really means for the first time.
The essence of organizational culture is: employee behavior patterns that constitute a key element helping the company succeed.
So the first-principle of organizational culture is actually: business nature determines organizational culture.
For example, ByteDance and Huawei are both companies with strong organizational capabilities, but if you swapped their organizational systems, both would likely collapse before long. They are at two extremes of the same spectrum: ByteDance emphasizes 'daring to be first,' Huawei emphasizes 'daring to be last.' One values innovation more, the other values efficiency more.
This has nothing to do with value judgments but is determined by business nature. When creating a new product, Huawei works on things like base stations and chips; if problems occur, recall costs could swallow a year's profits. ByteDance is different; it's typical short-cycle, short-chain business, capable of running dozens of versions in a week, fixing mistakes and re-releasing.
So ByteDance can encourage innovation, choose 'Context, not Control.' Huawei cannot. For Huawei, premature innovation might be a burden. What Huawei excels at is, after PMF (Product-Market Fit) appears in the market, surpassing and eventually crushing competitors through its organizational capabilities and resources.
Now back to Anthropic.
In AI competition, a core moat is enabling 'smart people to do dirty work.' Especially in the Coding and Agentic direction, on the surface, it's model capability competition; deeper down, it's actually engineering capability competition. It's not a problem that a few geniuses can solve with a flash of inspiration but involves a lot of dirty, fragmented, detailed systems engineering.
The most core barrier is data.
Past chat data was simple text data, but Coding and Agentic data are more complex—not just conversation logs but also the tasks themselves, environment setup, execution trajectories, and the entire evaluation and verification system.
All of this is dirty work, crucial if done well, but unlike publishing a paper or launching a new product, it doesn't become a personal highlight moment.
Based on feedback from exchanges with some researchers, a core problem at OpenAI today is its difficulty organizing hundreds of top-tier people to diligently work on data and do dirty work.
OpenAI hires top-of-the-hierarchy talent with great backgrounds and high aspirations. People naturally want to make their own bets, go from 0 to 1. As for cleaning up messes and supplementing data, few are willing.
OpenAI succeeded this way in the past; it indeed gained huge leads through some core paradigm breakthroughs. But as Yao Shunyu recently said in an interview: 'The era of individual heroism is over,' 'AI doesn't require much brains... The most important trait is reliability, meticulousness.'
At this point, one realizes Anthropic's low-ego, highly cohesive, mission-driven atmosphere's advantages become very pronounced.
It's said Anthropic co-founder Jared Kaplan leads the team daily in reviewing data personally; data cleaning is done extremely meticulously, something no other company achieves.
(This also explains a phenomenon: OpenAI's models are strongest on competition-level coding problems because such tasks are more a research problem, but often lag behind Anthropic on daily work agentic tasks because the latter is more an engineering problem, testing data, systems, and execution details.)
II. The Founding Team's Background
A company's values can be seen as part of the founder's values, like Jack Ma's martial arts style, Pony Ma's soft openness, Steve Jobs' aesthetic orientation, Ren Zhengfei's military discipline.
More accurately, founders' values often come from two parts: what the founder originally believes, and what they deeply disliked before.
The former determines what you want to become; the latter determines what you absolutely do not want to become again.
Anthropic clearly has both, and the shaping force of the latter might be even stronger. Let's briefly look at Dario's experience:
Dario first encountered AI at Baidu's AI Lab, where he first observed scaling laws and gradually became a firm believer. But after making breakthroughs at Baidu, internal struggles over control and resources erupted quickly, ultimately dissolving the team.
Dario later moved to OpenAI, deeply participating in advancing the GPT series. OpenAI once gave him 50%-60% of the company's total computing power, letting him lead the GPT-3 project.
Because Dario is someone with distinct values and personal opinions, his organizational philosophy differences with others at OpenAI began to surface.
For example, Greg Brockman once proposed a startling idea: in the future, AGI could be sold to nuclear powers on the UN Security Council. Dario almost resigned on the spot. In his view, this wasn't a business disagreement but a fundamental values issue.
Greg and Dario were at odds for years; Sam Altman mediated. Sam played to his strength: making different factions feel he was on their side. Short-term, this is balancing act; long-term, it's eroding trust. Later, comparing notes, people realized Sam's promises to Dario and Greg were completely different things.
Gradually, Dario formed a tight-knit circle within the company. Some called this small group 'the pandas' because Dario likes pandas. Their disagreements with OpenAI leadership on route choices, organizational governance, etc., grew larger, turning into serious political struggles.
Leadership even had a serious confrontation. Sam accused Dario and Daniela (Dario's sister, later an Anthropic co-founder) of organizing negative feedback about him behind his back; they denied it and called Sam's alleged source to confront him on the spot. The source said they knew nothing about it, then Sam turned around and denied making the accusation.
This made Dario and his sister lose trust completely; they argued on the spot.
There were many similar internal dramas. Ultimately, Dario escalated the conflicts to a moral trust crisis. He felt a company wielding such powerful technology must have sincere, trustworthy leaders. If those at the helm aren't honest, they are adding bricks to a dangerous direction.
Thus, Dario finally left OpenAI with some core colleagues from GPT-3 and founded today's Anthropic.
So, Anthropic's culture today isn't just because Dario is inherently like that; more importantly, he personally experienced political struggles at both Baidu and OpenAI. He knows how easily a group of smart, high-ego people can split due to resource competition and value differences. So they instinctively built Anthropic in the opposite direction:
Having seen how balancing acts erode trust, they emphasize authenticity and transparency;
Having seen intensified political struggles, they encourage bringing conflicts forward and talking them out early;
Having seen organizational collapse due to philosophical differences, they implement strict cultural screening;
Having seen superstar power struggles, they emphasize low ego and avoid hiring big names.
Much of Anthropic's organizational culture today seems like a reaction force from the experiences at Baidu and OpenAI.
03. Conclusion
To summarize, Anthropic and OpenAI are companies with quite different underlying tones. The former is an idealistic, mission-clear, highly cohesive cult-like organization; the latter is ambition-driven, multi-line expanding, constantly seeking the next big hit super-platform.
To see more clearly, we can list several core dimensions of both companies side by side:
However, although we discussed many of Anthropic's strengths, it's hard to conclude that one culture necessarily triumphs over another or predict the landscape three months from now. The AI world changes too fast, and OpenAI is now being underestimated by the market. For example:
• Coding is already an open hand; OpenAI could likely catch up. A clear trend now is developers migrating from Claude Code to Codex;
• Demand explosion far exceeds everyone's expectations; computing power is becoming the new deciding factor, and OpenAI locked in far more computing resources than Anthropic early on;
• OpenAI's culture of open exploration has its own huge advantages, and OpenAI is also always more aggressively exploring and betting on new paradigms; the next leap could flip the situation.
We can only say, looking back at the past three years from 2026, Anthropic has indeed left a noteworthy sample for the entire industry:
In the AI era, winning doesn't necessarily require greater ambition, more exploration, and stronger talent.
Sometimes, winning can also come from the opposite: fewer bets, lower ego, and a naive mission.
P.S. We are also curious about what organizational cultures and best practices other frontier AI companies are forming. We welcome friends with firsthand observations & thoughts to contact us via the details below for discussion!
Perhaps the next great AI company is, first and foremost, a new kind of organizational invention.
