First to Risk

Between $ and $

In September, I wrote about how Anthropic was preparing to go public (more on that here: "The Observer's Position: How Anthropic Is Preparing for a Trillion-Dollar IPO"), and I opened that piece with a caveat: the prospectus had been filed confidentially, and there was no public version.

On October 26, that public edition appeared on the website of the U.S. Securities and Exchange Commission. It contains more than 300 pages.

The cover states that Anthropic's price per share will be "between $ and $," with nothing between the dollar signs, and the prospectus doesn't specify either the underwriting commission or the amount the company stands to receive. A list of 22 underwriters, banks that will place the shares with investors and partner with Anthropic on setting the price, includes Goldman Sachs, Morgan Stanley, J.P. Morgan, Citigroup, and Deutsche Bank.

Reading the opening, you could mistake the prospectus for a marketing deck. The first line reads: "For the first time in history, intelligence is something humans can build." A few lines down, the text says AI will reshape the global economy more profoundly than industrialization, electricity, semiconductors, or the internet. Millions of "Virtual Collaborators" will work in parallel, it predicts, and the result will be something with no historical precedent: "a country of geniuses in a data center." Anthropic goes on to offer possibilities like "compressing decades of scientific progress into years."

Claude is already building the next Claude: writing code, running experiments, and checking safety. The prospectus calls this "closing the loop," one of the company's most significant structural advantages. The risk factors section notes more soberly that Anthropic assesses its own pace mostly in hindsight, so it may notice a significant acceleration only after it has already happened.

As I explained in my previous article, five trustees, who hold no stake in the company, will elect a majority of the board. The prospectus adds a new detail: the seven founders will share a single Class F share. That share carries no economic rights, but it confers 50.1 percent of the votes at shareholder meetings. Anthropic itself warns that this could leave holders of common stock with no meaningful say over major decisions.

After establishing the structure, the prospectus lets the lawyers take over. Nearly a hundred pages of risk factors follow.

ANT

On October 24, Thariq Shihipar, one of the engineers behind Claude Code, announced on X that he had left Anthropic. Shihipar had spent more than a year working on Claude Code and watched it grow bigger than he had ever imagined. Now, he wrote, it was time to rest a little and try something new. He believes deeply in the company's mission and in the idea of a "race to the top," a phrase CEO Dario Amodei uses for Anthropic's ambition to show that AI can be built responsibly and to raise the bar for the entire industry.

In the same post, Shihipar wrote that he had really loved working at "Ant," his name for the company. Two days later, those same three letters appeared on the cover of the prospectus: ANT will be the company's ticker, the symbol under which its shares will trade on Nasdaq.

At about the same time, Anthony Morris, the creator of the Claude Code desktop app, announced his own departure on X: "I created Claude Code on desktop because I believe everyone can be a builder." He called the bug reports and feature requests "a gift that impacted millions of lives," and added that "the product wouldn't be what it is without you."

In their farewell posts, both thanked their colleagues. Neither explained why he was leaving.

Risk factors

Much of what happened at Anthropic in 2026 made it into the prospectus. Shihipar and Morris left without giving reasons, but others had stated theirs earlier.

Safety team leader Mrinank Sharma, for instance, departed in February. In a letter to colleagues, he wrote that the "world is in peril" and that it is very hard, in practice, to let values govern actions. Jacob Coxon's September exit followed three years of training models, first at OpenAI and then at Anthropic. On his way out, he said that AI companies are "gambling with our lives": Anthropic understands the risks but is "locked in a race to get there first." Evan Hubinger, Anthropic's alignment science lead, backed Coxon: "Jacob is correct here — we really do earnestly believe AI could kill all humans!"

The company acknowledges that employees may become disillusioned and lose confidence in leadership if they conclude that Anthropic is departing from its values, and that public criticism could make hiring and retention harder and more expensive. It also acknowledges that concerns about AI's catastrophic risks exist inside the company itself, as Hubinger's comment shows; the doubts aren't only coming from former employees.

Similarly, the document notes incidents in which bugs in the cybersecurity test environment (the "sandbox") let models evade their restrictions and get onto the internet. One such model reached a real company's database, another published malware in the open, and a third broke into someone else's application. Even more dramatically, another model used the Philadelphia Police Department's website to send a false tip about an unsolved homicide (the message landed in a spam folder, so it didn't influence the investigation, and Anthropic itself didn't discover the incident for more than two months).

Anthropic confirms that its models gained access to outside systems and states it notified those affected and didn't find any signs that the models tried to evade control or collude with one another. It also warns that the models could take harmful actions on their own, without direct human instruction, and that such actions may become harder to monitor.

Internal evaluations now bar models from the internet, and Anthropic will restore their access only once it is confident that monitoring reliably catches problematic behavior. Another price of safety: training for a new model has already been delayed by more than a month to tighten protection against model-weight theft, which could put Anthropic at a disadvantage against competitors that face no such constraints.

Price

All year, the company's valuation kept climbing: $965 billion in May after a funding round, then $1.2 trillion in July on the secondary market. By fall, investors were expecting $2 trillion or more. Then, in mid-October, at a closed-door meeting in San Francisco, Anthropic disclosed its third-quarter revenue — $16.8 billion.

That's more than triple the figure for all of 2025, yet some investors came away disappointed. It wasn't the profits that let them down: in 2025 Anthropic lost more than $8 billion from operations, and now it has posted an adjusted profit for the second quarter in a row. The problem was the trajectory: revenue more than doubled from the first quarter to the second, but grew by only 45 percent from the second quarter to the third. After the October disclosure, the expected IPO valuation fell from $2 trillion to $1.5 trillion.

A week before the prospectus was published, traders on Kalshi, an American prediction market, gave Anthropic 75 percent odds of going public this year. No other large private company's figure even reached 50 percent.

Anthropic's main competitor isn't even in the running. In September, OpenAI CEO Sam Altman told Fortune that, "given everything happening with safety, right now would be an ill-advised moment to go public." Asked about the whole year, he added: "I would say not 2026."

Around the same time, Amodei published an essay on his personal website arguing that "we must slow the pace at which we improve the capabilities of AI models." Altman replied on X that he agreed. Anthropic's plans, however, did not change.

The company is taking a risk — and taking it first. It'll fill in the share price later.