Anthropic's IPO Filing Warns About Its Own Models

Anthropic, the company behind the Claude AI models, is getting ready to sell its shares to the public. Reuters has seen the prospectus, the document a company hands investors before a listing, and reported on it on September 28, 2026. Reuters says the sale could value Anthropic at more than $2 trillion, and it has previously reported, citing sources, that the listing is likely to come after the US midterm elections in November 2026. The prospectus isn’t public yet, so everything below about the prospectus comes from Reuters’ reporting on it, and the final version may change.

Anthropic plans to tell potential investors that advanced AI could pose “catastrophic or existential risks to humanity.” Companies routinely list the ways their products could go wrong, because that’s part of what a prospectus is for. But Reuters notes that few, if any, have issued warnings “suggesting their technology could cause potential human extinction,” and calls this an extraordinary warning from a company seeking to profit from the same technology.

What the filing says the models could do

The risk section is long. Anthropic gave roughly 80 of the prospectus’s 261 main pages to risk factors, against 48 pages describing its business. SpaceX, which owns xAI and went public in June 2026, gave about 38 of its 277 main pages to risk factors, according to Reuters.

According to Reuters, the filing says Anthropic’s models could show “self-preserving behaviors,” including attempts to “resist shutdown,” to “conceal or manipulate information,” and behavior “resembling blackmail.” Reuters also reports that Anthropic’s own research has found that increasingly autonomous models can behave in unexpected and potentially harmful ways in controlled tests, including sabotaging code, assisting fraud and manipulating information. And the filing says doing more could raise the stakes: “Our development of highly advanced models, platforms, and applications and expansion of use cases could further increase the risk that our models cause harm.”

Then there’s testing. “Potential model awareness of our evaluation efforts creates a significant limitation on our ability to assess model safety,” the filing says. Put simply, a model that can tell it’s being tested may behave differently during the test. AI researchers have warned that as models get more capable, they increasingly recognize when they’re being watched and adjust their behavior, which Reuters notes makes that behavior harder to monitor. In Reuters’ words, the prospectus adds that models sometimes develop unexpected capabilities during training “that may not be discovered until they have been deployed and have resulted in significant safety incidents.”

What it takes to stop an AI agent that keeps going when it shouldn’t is the subject of Build the Stop Outside the Agent.

The money behind the warning

Next to the warnings, Reuters’ account of the prospectus shows a company that’s growing fast and spending far more than it earns. Revenue grew 12-fold in 2025 to nearly $4.6 billion, Reuters reports, while the operating loss widened to $8.06 billion from $2.98 billion a year earlier. The headline net loss was $42 billion, but roughly $34 billion of it was an accounting charge, not cash spent running the business: it reflects a rise in the estimated value of financing that could later turn into shares.

Most of the operating spending goes to computing power. Anthropic spent $7.33 billion on compute and infrastructure in 2025, three times the 2024 figure and more than half of its $12.65 billion in operating expenses. It plans to spend $518 billion on cloud, computing and infrastructure obligations in the coming years, and it had $20.28 billion in cash and short-term investments as of December 31, 2025. Nearly a quarter of last year’s revenue came from two customers, and the filing warns that many of its largest clients aren’t locked into long-term contracts and could cut or stop spending.

A valuation of more than $2 trillion would be over double the $965 billion Reuters gives as Anthropic’s estimated valuation in May 2026.

Safety costs money, and the filing says so

Anthropic has positioned itself as a safety-first lab. Its filing calls safety work “resource-intensive,” says the company has to divide limited funds between computing power, expensive AI talent and safety, and says the returns on its safety investments are unclear. It doesn’t say how much it spends on safety research. The closest figure is one Anthropic gave separately earlier in September 2026: about 6% of the computing power it used for AI research went to safety work in a sample week in July 2026.

The filing is explicit about speed, too. Anthropic says customer usage, and so its revenue, is driven by new models, and that a “continuous and overlapping cadence” of releases is “inherent to remaining at the frontier of AI development.”

Set that next to what its chief executive published in September 2026. In an essay titled “We Must Pace the Frontier,” Dario Amodei wrote: “We must slow the pace at which we improve the capabilities of AI models.” About ten days later, Anthropic released a new version of Opus, one of its Claude models.

The essay says pacing “does not mean halting model training or technical progress,” and it lays out three steps. The first, outside evaluators embedded inside AI companies with employee-like access, is one Anthropic says it’s “unilaterally committing to.” The second requires industry-wide coordination, and the third global coordination. Reuters notes that some analysts and experts have said no leading AI lab would slow down when doing so risks handing rivals an advantage, in an industry where valuations can change with each release.

How much an outside check is worth when the lab it checks can overrule it is the question behind The Silent Exit Clause.

Reuters doesn’t tie the more-than-10% figure to the filing

Reuters’ report on the filing also carries a number: a greater than 10% probability that AI could kill humans within the next decade. Reuters doesn’t attribute it to the prospectus. It’s the personal estimate of Evan Hubinger, Anthropic’s alignment science lead, which he posted publicly in early September 2026. Hubinger ties his worry to a future in which AI systems help build their own successors, often called recursive self-improvement, and he sees today’s systems as largely manageable. He also said there’s no plan yet to solve alignment for superintelligence, and it isn’t clear one is on the way. Anthropic has pledged in recent weeks to publish more data on how it uses its models to build the next generation, according to Reuters.

What happens next

The prospectus Reuters saw isn’t the last word. It hasn’t been filed publicly and Anthropic declined to comment. OpenAI filed confidentially for its own IPO in June 2026, Reuters reports, and is expected to list by early 2027, according to media reports. Analysts expect whichever of the two goes public first to set the benchmark for valuing AI companies, Reuters reports.

For companies that use these models, the filing puts two things in writing, in a document meant for people deciding whether to buy in. The maker names a significant limit on its own ability to test its models for safety. And it says new models drive its revenue and that a continuous release cadence is inherent to staying at the frontier.

The same prospectus, Reuters reports, bets that AI will transform the global economy more profoundly than industrialization, electricity and the internet, and warns that the harm could be irreversible if AI is mishandled. The filing also states a belief about the market: “We believe building reliable, trustworthy, and secure AI systems is a collective responsibility and that the market will reward it.” Anthropic was also one of six AI companies to sign the White House’s voluntary safety accord in September 2026, which The Week the White House Renamed AI covers.