Anthropic Launches Its Most Powerful Agent Model, Claude Sonnet 5

Author | Zhou Zhou

Source | Tech Walker

[Editor's Note] As one of the world's large language models, Claude is very popular and useful, but who knew that the company behind it, Anthropic, had so many stories?

This article tells the story of several founders who left OpenAI, bringing with them a philosophy that "AI must be ethical," to found the company Anthropic, and cultivate (as they say) "the most upright citizen in the AI world," Claude.

Interestingly, this article also reflects the two sides of AI and humanity. The seemingly gentle and cute Claude actually knows how to "pretend to be good" — tests show it can pretend to be obedient while secretly doing "bad things." Co-founder Dario Amodei worries about AI destroying the world on one hand, while envisioning an AI heaven on the other. This contradictory attitude actually reflects the industry's entanglement.

Of course, for large model startups, there is also something to learn about how to find a balance between making money and safety.

Departure and Restart

Whenever Anthropic founder and CEO Dario Amodei talks about AI, he gets energized. He jumps up from his seat, rushes to the whiteboard to draw diagrams, and the curves he draws are always steeply upward, showing how machine intelligence rushes toward infinite possibilities. When speaking, he unconsciously messes with his fluffy curly hair, as if soothing his overheating brain.

Amodei firmly believes his company is different — they are not just "building AI," but creating "powerful AI" that never does evil, a "guide" that can lead humanity into an ideal world.

However, at Anthropic, the real protagonist is not Amodei, but a large language model — Claude. Nowadays, it stands shoulder to shoulder with large models like OpenAI's ChatGPT, Google's Gemini, and Matrix's DeepSeek.

Previously at the Davos World Economic Forum, Amodei boldly predicted: within two years, Claude and its peers will surpass humans in all thinking tasks. But upon returning to the company, he and Claude faced a crisis. The popularity of DeepSeek R1 during the Spring Festival shocked the industry and shook the industry consensus — does building cutting-edge AI really cost billions of dollars?

Amodei is a representative figure of the "Big Compute Theory (Big Blob of Compute)." When he worked at OpenAI, he proposed: if you feed AI massive amounts of raw data, paired with powerful computing power, it can accelerate the birth of super AI. This theory has become an industry standard, and is why only wealthy companies can play this game.

Now, DeepSeek succeeded by taking an unconventional path. But Amodei is not worried. He frankly admitted: "This precisely proves that intelligence is more valuable! If every dollar can exchange for more intelligence, you will only want to invest more money!" This also explains why even after DeepSeek, OpenAI and Microsoft still announced plans to invest hundreds of billions of dollars in data centers.

In fact, what Amodei truly cares about is how to safely reach AGI (Artificial General Intelligence). It is this question that led him and six other founders to leave OpenAI initially — they did not trust Sam Altman (OpenAI founder) to handle this matter. At Anthropic, they are racing against time to set global standards for all future AI models, ensuring AI truly helps rather than destroys humanity. Amodei calls this the "Race to the Top."

Walking into the Anthropic office, you will find Claude everywhere and nowhere. It has no physical form, yet participates in every aspect of the company. From the day it was born, Claude has been helping engineers create a better Claude. If Amodei's vision comes true, when humans enter the new AI era, Claude will become the guide.

But this raises a wonderful question: Can we really trust that Claude will always do good?

Among Anthropic's founding team, there is a special member — Amodei's younger sister, Daniela. In the 1970s, their parents, Elena Engel and Riccardo Amodei, moved from Italy to San Francisco. Amodei was born in 1983, and his sister Daniela is four years younger. Father Riccardo was a leather craftsman from a small town near Elba Island, who fell ill when the children were very young and passed away when they were adults. Mother was a Chicago-born American Jew who worked as a project manager in a library.

Anthropic Co-founder Daniela

Amodei has been with numbers since he could speak. When children his age were still holding blankets, he was already playing with calculators. As he grew older, he became obsessed with mathematics. "I was just addicted to manipulating mathematical objects, understanding the world in a quantitative way," he said. In high school, he naturally buried himself in math and physics courses. His sister Daniela, however, studied humanities and music, and even received a classical flute scholarship. But Daniela said that she and her brother both have humanistic sentiments. When they were young, they often played games about saving the world.

When going to college, Amodei originally wanted to become a theoretical physicist, but soon realized this field was too far from the real world. "I strongly felt that I wanted to do something that could promote social progress and help humanity," he said. Until a professor in the physics department was researching the human brain, which sparked his interest. He also began reading works by famous American futurist Ray Kurzweil about "non-linear technological leaps." Later, Amodei completed an award-winning computational biology doctoral dissertation at Princeton.

In 2014, he joined Baidu's US research laboratory. Under the guidance of Andrew Ng, Amodei began to understand how a massive increase in computing power and data volume could produce qualitative leaps. At that time, some people were already worried about the risks these systems posed to humans. He was initially skeptical, but after moving to Google in 2015, he changed his mind. "Before, I thought, we are not building those kinds of systems, what could happen? But now we are indeed building these systems."

Around that time, Sam Altman found Amodei and talked about a startup dedicated to building AGI safely and openly. Amodei attended the later legendary Rosewood Hotel dinner, where Sam Altman and Elon Musk pitched their ideas to venture capitalists, tech executives, and AI researchers. "I was not convinced," Amodei said, "I became even more skeptical. Because their goals were not clear to me, it felt more like a gathering of celebrity tech investors and entrepreneurs, rather than a gathering of AI researchers."

A few months later, OpenAI was established as a non-profit company, stating its goal was to advance AI development to make it "most likely to benefit all of humanity, unconstrained by the need to create financial returns." Amodei was attracted by the top talent drawn to the company, including some of his old colleagues from Google Brain, so he also joined OpenAI's bold experiment.

At OpenAI, Amodei refined his ideas. It was at this time that he wrote the "Big Compute" paper elaborating on the scaling theory, the implications of which seemed even more terrifying than before. "My first thought was, oh my god, will systems smarter than humans find ways to disrupt the balance?" Shortly after, an engineer named Alec Radford applied the big compute concept to a breakthrough AI technology called "transformer," and GPT-1 was born.

Around that time, sister Daniela Amodei also joined OpenAI. Her career was quite tortuous: graduated with a degree in English from college, a super fan of Joan Didion, worked overseas in NGOs and government for many years. Finally returned to the Bay Area to become an early employee at Stripe. Looking back, the development of GPT-2 might have been the turning point for her and her brother. Daniela managed that team. The model's coherent, paragraph-style answers seemed to be early signs of superintelligence.

Witnessing all this made Amodei both excited and fearful. He said, "We hold one of the world's craziest secrets, which will determine the fate of nations."

Amodei urged OpenAI colleagues not to release the full model immediately. They agreed and publicly released a smaller, capability-limited version in February 2019. They explained in their blog that this limitation was to set an example of "responsible behavior" in the AI field. Amodei said, "I'm not sure if this model is dangerous, but my feeling is that we should do something to mark it — clearly indicate that these models might be dangerous." A few months later, OpenAI released the full model.

Discussions about responsibility began to shift. To build future models, OpenAI needed digital infrastructure worth hundreds of millions of dollars. To obtain funding, the company expanded its cooperation with Microsoft. OpenAI established a for-profit subsidiary, and soon almost all employees were placed under it. The company began to take on the appearance of a typical Silicon Valley tech company.

Some employees began to worry about the company's direction. Pursuing profit was not the problem, but they felt OpenAI was not prioritizing safety as they hoped. Among them, unsurprisingly, was Amodei. He said, "One reason I was frustrated was that as these issues became more serious, the company started moving in the opposite direction." He expressed his concerns to Sam Altman, who would listen seriously and agree, but nothing changed afterwards. (OpenAI chose not to comment on this matter, but its stance is that safety has always been the top priority.) Gradually, these doubters found each other and shared their concerns. As one of them said, they began to ask themselves: Are they really working for the "good guys"?

Amodei recalled that when he told Sam Altman he was resigning, the CEO repeatedly tried to retain him, but he realized he should have left earlier.

At the end of 2020, he and six OpenAI employees, including his sister Daniela, resigned to found their own company — Anthropic.

The Birth of Anthropic

Talking about the birth of Anthropic, Daniela remembers a photo taken in January 2021, where this group of "defectors" gathered for the first time under a large tent in Amodei's backyard. Former Google CEO Eric Schmidt was also present, listening to their startup concept. Everyone was wearing COVID masks, and it was pouring rain outside. Within weeks, more than ten more people jumped from OpenAI to this new competitor.

Schmidt did invest in Anthropic, but the initial $124 million in funds mainly came from sources related to the "Effective Altruism" (EA) movement. The concept of EA is that successful people should use their income for charity. In reality, EA enthusiasts are keen on specific causes, including animal rights, climate change, and the potential threats AI may pose to humanity.

The main investor in Anthropic's seed round was EA supporter Jaan Tallinn, an Estonian engineer who earned billions by creating Skype and Kazaa, and invested funds and energy into a series of AI safety organizations. In the second round of financing (the fund pool increased to over $500 million), the main investors were EA advocate (now a convicted criminal) Sam Bankman-Fried and his business partner Caroline Ellison. (Bankman-Fried's shares were sold in 2024) Another early investor was Facebook co-founder Dustin Moskovitz, who is also a strong supporter of EA.

These investments led to a strange, years-long "romantic comedy-style" interaction between Anthropic and EA. Asked about this, Daniela said: "I am not an Effective Altruism expert, I do not agree with that statement, my impression is that this term is a bit outdated." However, her husband Holden Karnofsky co-founded one of EA's most famous charities, he publicly talks about AI safety, and joined Anthropic in January 2025.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.