Anthropic Files for IPO: Valuation Falls to About $1.5 Trillion, Safety Commitments Give Way to Expansion

Wen | Baobian, Author | Zhan Fangxu, Editor | Xing Yun

On October 26, Anthropic publicly filed its prospectus, planning to list on Nasdaq under the ticker "ANT." According to Reuters, due to slowing growth in the third quarter, the valuation under discussion among investors has dropped from about $2 trillion to about $1.5 trillion.

As the "money-making" top student among AI large-model companies, Anthropic's revenue scale has achieved a great leap over the past three years: revenue was $41 million in 2023, reached $386 million in 2024, and was $4.59 billion in 2025, growing about 10-fold each year.

The prospectus shows that in the first half of 2026, the company's revenue reached $16.286 billion, 3.5 times that of the full year 2025.

On the books, the company had a net loss of $80.7 billion in the first half, but $72.4 billion of that was an accounting expense from the appreciation of convertible bonds; on an operating basis, the loss has narrowed to $693 million, and operating cash flow has already turned positive. At the final round of private financing in May this year, the company was valued at $965 billion.

In the more than 300-page prospectus, Anthropic depicts the "age of miracles" that AI will bring, while devoting large sections to warning that AI may threaten human survival.

Anthropic's seven founders left OpenAI back then to build a kind of Eden, but every company born of idealism still has to enter the market—for example, Google removed "Don't be evil" from the opening of its code of conduct, and OpenAI set up a "capped-profit" entity beneath its nonprofit organization.

Five years later, it is Anthropic's turn. Now, for the sake of going public and the pace of expansion, the original idealism is giving way.

1. The original idealism

Anthropic's founding team almost entirely came from OpenAI. In 2019, OpenAI adjusted from a nonprofit structure to a "capped-profit" model and brought in a $1 billion investment from Microsoft. At that time, some core OpenAI members began to worry that the company's technology roadmap would tilt toward commercial returns, and that AI capabilities would further concentrate in the hands of a few institutions.

At the end of 2020, Dario Amodei, then vice president of research at OpenAI, and several colleagues left one after another, including his sister Daniela Amodei. In 2021, the seven co-founded Anthropic.

Co-founder and CEO Dario Amodei later explained that the team's judgment at the time was that they believed scale would bring capability, but making models bigger alone was not enough—alignment and safety were also needed. Growth in compute would continue to improve model capabilities, but models would not automatically learn to comply with human intentions and boundaries. Preventing AI from outputting harmful content and deviating from design goals—these tasks must be addressed with dedicated resources in R&D.

Perhaps for this reason, they named the company Anthropic, meaning "related to humans." Daniela explained publicly that the name was meant to express the hope that AI development would always put people at the center.

This judgment about safety was also written into the structure the founders designed for the new company. For example, from its first day, Anthropic was registered as a public benefit corporation, and its board of directors is legally required to weigh both the company's mission and shareholder returns, not just make money for shareholders. Anthropic has also repeatedly stated its public benefit goal on various occasions: "to responsibly develop and maintain advanced AI for the long-term benefit of humanity."

After the completion of its Series C financing in 2023, the company formally established a governance arrangement with a long-term benefit trust, granting the trust the right to appoint some directors and adding a check between shareholders' economic interests and the public interest mission.

The prospectus published this time also fully discloses the details of this design. After listing, the trust in principle has the right to appoint and remove four of the seven board seats. It is worth noting that the trust does not participate in the company's day-to-day operations or strategy-making.

After listing, an entity jointly controlled by the seven founders will hold the sole share of Class F common stock, with a fixed 50.1% voting right on general matters submitted to shareholders for vote. This ratio will not be diluted by ordinary share issuance, and it can also decide the appointment of another three board seats.

Thus the company has formed a set of separated-power arrangements. The founders hold control over general shareholder voting, while the trust checks company decisions through its right to appoint a majority of board seats.

The founders' letter in the prospectus discloses the team situation: all seven founders are still employed to this day, and more than three-quarters of the earliest 100 employees are still with the company.

Beyond the company structure, the safety-first approach is also reflected in the pace of operations: Anthropic decided to do research first and products later. Daniela, one of the founders, recalled that in the first year after the company was founded, almost all of it was spent building the company, raising funds, and training models from scratch, and most of the early team was doing safety research. To this day, the prospectus still defines the company as "first and foremost a research laboratory."

In the summer of 2022, after Claude finished training, the company did not release it, fearing it would trigger a race to develop stronger models. About three months later, OpenAI launched ChatGPT, and the race began anyway.

2. The safety bottom line is being lowered

From a business model perspective, "safety" is an important selling point for Anthropic.

The company also said that the values followed by its public models enhance customer trust and have commercial value. According to the new version of the "constitution" released by Anthropic at the end of January 2026, the four core values are prioritized as follows: broad safety, broad ethical conduct, compliance with company guidelines, and being genuinely helpful.

For enterprise buyers who value safety, a more cautious model is itself a reason to purchase.

Prospectus information shows that all of the top ten companies in the Fortune U.S. 500 are Anthropic customers with annual spending exceeding $100,000, and among the 500 companies this proportion has also exceeded half.

It is reported that Anthropic's main revenue model is "paying for intelligence," measured by the number of tokens customers invoke and priced according to the selected model. In the first half of 2026, this usage-based revenue was $13.1 billion, accounting for about 80% of total revenue.

As of September 30, 2026, more than 11,900 customers had spent more than $100,000 at Anthropic over the past 12 months; a year earlier, there were only more than 950. As of the end of September, the company's net revenue retention rate was about 900%, meaning that the customer base from a year earlier now spends a total of nine times what it did then.

However, in the increasingly fierce AI competition, Anthropic's safety bottom line is loosening.

Previously, Anthropic did not use consumer conversations to train models by default, which was also a selling point of its emphasis on privacy protection. In August 2025, the company revised the relevant terms: if free, Pro, and Max users choose to agree, new conversations will be used for training, and the data retention period will be extended from 30 days to 5 years.

In February 2026, Anthropic revised its public "Responsible Scaling Policy," abandoning a public commitment in the 2023 version: not to train or deploy models that could cause catastrophic harm. Under the original design, once safety measures could not keep up, the company would need to pause training. This commitment was precisely the core that the company used in 2023 to prove it was different from its peers.

The company's explanation is that it is difficult to determine whether model capabilities cross the risk threshold, government legislation is progressing slowly, and risk prevention requires joint action across the entire industry. Co-founder and Chief Science Officer Jared Kaplan put it more directly: with competitors advancing rapidly, "our stopping training AI models actually doesn't help anyone."

In 2022, Anthropic's founding team would rather miss the first-mover opportunity than trigger a race.

Now, the race is a foregone conclusion and unavoidable, and Anthropic's release pace has clearly accelerated accordingly. In April 2026, the company opened a gated preview of Mythos; in June it launched Fable 5 and Mythos 5; in September it released Fable 5.1 and Opus 5.5 at the same time.

Rapid releases also pay off. In the fourth quarter of 2025, Anthropic's quarter-over-quarter revenue growth rate was 48%, and by the first quarter of 2026 it had rapidly jumped to 132%. Behind this, the release of Opus 4.5 was one of the important drivers.

The prospectus also shows that model versions released in the six months ending September 30, 2026 contributed more than 80% of revenue in the third quarter.

The commercial explosive power of new models is evident, but this also means that the company's revenue is highly dependent on continuous model iteration.

While the pace has accelerated, the prospectus also disclosed loopholes in safety enforcement. The company admitted that its safety classifiers and other controls have experienced failures, misconfigurations, and even being accidentally turned off, and that the protections of its strongest models have also been bypassed. In safety evaluations, the company observed that models behave differently when they perceive they are being tested, so evaluation results may be incomplete.

At the end of 2020, Amodei and others disagreed with OpenAI on safety issues and eventually left to found Anthropic; now, the same criticism is beginning to appear inside Anthropic.

According to Fortune, in September of this year, Jacob Coxon, who had worked on model training research at both OpenAI and Anthropic, resigned and publicly criticized both companies for failing to act responsibly.

The prospectus also lists such situations as risks: employees who joined because they identified with the company's safety commitments may suffer low morale and lose trust in management if they believe the company has not lived up to its stated values; public criticism by current or former employees of the company's safety practices, or the departure of safety-related personnel, may reinforce this perception and affect the company's ability to recruit talent.

3. The industry slows down, but the IPO must be fast

To outsiders, Anthropic is the most principled among the leading large-model companies.

At least since January 2026, Amodei has been publicly talking about the industry slowing down. In his long article "The Adolescence of Technology," he proposed that to address the risks brought by AI, the industry should seek a slight slowdown. In September, he published the article "We Must Set a Speed Limit for the Frontier," formally calling on the entire industry to slow the pace of improving frontier AI capabilities to buy time for safety research.

Regarding the phrase "industry slowdown," the prospectus specifically explains that this call is aimed at the entire industry working with governments to set standards, and does not describe the company's own development and release timetable.

Meanwhile, Anthropic's IPO process has been accelerating. In December 2025, the company said it had not yet decided when, or even whether, to go public; six months later, it confidentially filed an IPO application with the U.S. Securities and Exchange Commission. Just one day before Amodei published his long article calling on the entire industry to slow down, Reuters reported that Nvidia was in talks to invest up to $10 billion as a cornerstone investor in Anthropic's IPO.

The Wall Street Journal once published an article saying that Amodei called on the entire industry to slow down, yet Anthropic itself has not slowed down, and is caught in a "prisoner's dilemma." And if the leading labs were to slow down in coordination, Anthropic, which is already in a leading position, would likely benefit as well.

Objectively speaking, Anthropic is a public benefit corporation, but it is also an absolute giant of a company, so in practical terms it indeed needs an enormous amount of capital to sustain its operations.

And the company's bottom line on funding sources is also adjusting along with its financing needs.

In March 2024, according to CNBC, Anthropic's founders had excluded Saudi investors from buying its shares on national security grounds. But according to a report by the U.S. tech media outlet Wired, in July 2025, Dario Amodei said in an internal letter that the company would seek investment from the UAE and Qatar, on the grounds that Middle Eastern capital is enormous in scale and would help the company stay ahead. He acknowledged that this might benefit "dictators," but believed that completely ruling out such a situation would make it difficult for the company to operate. Two months later, the Qatar Investment Authority participated in Anthropic's Series F financing.

Information in the prospectus shows that Anthropic has signed long-term contracts with multiple compute suppliers. For example, by July 2033 the company must spend at least about $114.5 billion with Google, by April 2036 at least $110 billion with Amazon, with any shortfall to be made up; and by May 2033 at least $44.9 billion with Microsoft, regardless of whether it is actually used.

Based on the data disclosed in the prospectus, Anthropic's future compute commitments total roughly $540 billion to $570 billion, of which about $417 billion is contractually agreed minimum payment obligations.

This scale of long-term spending is far higher than the amount of Anthropic's largest single private financing round to date. In May 2026, the company announced the completion of a $65 billion Series H financing, an amount equivalent to about 12% of the above estimate of compute spending. As of June 30, 2026, the company held cash, cash equivalents and short-term investments totaling about $131.7 billion.

Although these compute expenditures will occur gradually over many years, the company still needs to rely on continuously growing operating revenue and subsequent financing to support its enormous investment. Going public can also open up equity and debt financing channels in the public market for the company.

Five years ago, seven founders left OpenAI over disagreements on safety philosophy and founded Anthropic. Today, this company stands at a crossroads similar to that of OpenAI in the past: carrying more than $500 billion in compute commitments, about to face the public market's interrogation of growth, and also beginning to face employees leaving in the name of safety.

It has never stopped talking about safety, and it has indeed paid a price for it. But the wording in the prospectus also shows that when safety ideals and commercial growth meet head-on, ideals need to give way. In the business world, there is no Eden.