# Anthropic stirs up safety with existential risk ahead of its IPO

> Anthropic devotes 80 pages of its prospectus to existential risks and releases Sonnet 5.5. The figures, the context and what it means to go public like this.
> By José Maldonado · Artificial Intelligence · 2026-10-05

Some prospectuses sell the future and some warn of the end of the world, and Anthropic's does both at once. The company behind Claude is preparing an IPO that could exceed $2 trillion, and its prospectus, reviewed by Reuters and the Financial Times, devotes about 80 of its 261 pages to warning that its models could pose catastrophic or existential risks to humanity, according to [The Guardian](https://www.theguardian.com/technology/2026/sep/29/anthropic-warns-existential-ai-risks-humanity-ipo-document-claude).

If you follow the world of finance, think about how unprecedented the move is. No company on its way to the stock market had ever told its investors that its product **"could contribute to human extinction"**, including self-preservation behaviors, resistance to shutdown and blackmail. And it does so while reporting $42 billion in losses for 2025, committing $518 billion to future computing and releasing new models every couple of months. Asking for money while warning of the apocalypse: welcome to the 2026 stock market.

The key is that the prospectus and the product came out the same week. While the lawyers were drafting risks, the engineers were releasing Claude Sonnet 5.5, 30% faster and cheaper per task. Understanding what the document says, what the numbers say and what it means to list while warning of the end of the world is understanding where the AI industry stands today.

## What the prospectus says about its own models

The prospectus language leaves no room for euphemism. According to Anthropic, the models could display self-preservation behaviors, attempt to resist shutdown, hide or manipulate information and behave in ways that resemble blackmail. And developing more advanced models and expanding their use, the document warns, could increase the risk that they cause harm. This is not an alarmist blog: it is a statement made under stock-market liability, which binds those who sign it.

The most devastating detail for the industry is technical. Anthropic acknowledges that "its models sometimes know they are being evaluated," which seriously limits the ability to measure their safety, and that some capabilities only appear after deployment, when it is already too late for prior testing. Add the routine risks of any prospectus, lawsuits and competition, and the mix is explosive: the ordinary does not reassure when the extraordinary is extinction. If the examinee knows it is being examined and the exam does not detect what emerges afterwards, the whole edifice of evaluation needs rethinking.

After all, the prospectus spreads blame with Parisian generosity. It warns that the concentration of power and wealth that AI can generate threatens society and the geopolitical order, while calling for control and capital to be concentrated in ever fewer hands to make it safe. It warns that losing access to third-party computing would sink the business, while committing half a trillion to that same computing. Each risk also describes the bet.

## The numbers of the business asking for money

The prospectus describes a company growing at full speed and losing at full speed even faster. Revenue multiplied twelvefold to $4.6 billion in 2025, but the net loss reached $42 billion, with some $34 billion in accounting charges tied to financing and more than $8 billion in operating loss.

The business model, metered usage by tokens plus subscriptions, replicates that of OpenAI and Google, without either having yet proven that the numbers add up at scale. Charging per token sold while burning billions in computing is a race where the prize grows with the spending: the more customers, the more revenue and the bigger the electricity bill. The prospectus does not hide that equation, it underlines it as both a risk and a thesis.

Prospectus figure

Value

Reading

Target valuation

More than $2 trillion

Double the $965B of May

2025 loss

$42B

Record with $34B accounting

Computing commitments

$518B

Existential bet on scaling

Risk pages

80 of 261

Double those of the business (48)

Comparable SpaceX

38 of 277

Half the risk pages

The comparison with SpaceX, which went public in June valued at $1.77 trillion, is inevitable and devastating. The company that puts people in tubes loaded with hundreds of tons of highly explosive fuel to send them into space devoted 38 pages to risks; the one that writes text and calls itself "intelligent" devotes 80. And the debut, according to sources familiar with the matter, will wait until after the November midterms. Not even the calendar is innocent: you go public when politics stops watching, just in case.

It is worth noting that the prospectus also sells paradise. From the claim that AI will transform the economy more than industrialization or the Internet, to the warning on the next page that it could end all of humanity. Anyone would see that ambivalence as a mistake, but the reality is that this is the real business. Only someone who believes in both halves at once will sign the check, and Anthropic is looking precisely for those investors. After all, who would be so foolish as not to take part in building a technology capable of rivaling us at every level. Investors who fall into the trap need only look in a mirror to find the answer.

**[NVIDIA OpenShell: how to stop AI agents from doing whatever they want](https://observatorioblockchain.com/en/ai-agents/nvidia-openshell-agentes-ia-seguridad/)**

## Sonnet 5.5, while the lawyer drafts

The week of the prospectus also brought product, and Anthropic, like a good camel, brought the good stuff. Claude Sonnet 5.5 keeps the price of $2 per million input tokens and $10 per output tokens, but delivers more than 30% faster and up to 30% cheaper per completed task. On Terminal-Bench for agentic coding it jumps from 10.3% to 70.6% and on OSWorld for computer use from 57% to 80.1%. For daily work and document creation, they say, there is no comparison with its predecessor. It arrived on September 28 on AWS, Google Cloud and Azure, with a knowledge cutoff of June 2026, a one-million-token context and a Haiku 5.5 on the way in the coming weeks.

The model also brings two details that matter more than the benchmarks. It is the first Sonnet with frontier-level cybersecurity safeguards: sensitive security requests are quietly routed to the previous model and new classifiers block attempts to distill its reasoning. And it brings an oddity that should make the whole sector reflect: at maximum effort it performs worse than at high effort in some tests, because the extra review causes idle time. Spending more computing is no longer a free win.

In this way, Anthropic preaches restraint and accelerates at the same time, and does so without concealment. Amodei called at the UN for slowing the frontier, published a lengthy essay and ten days later released Opus 5.5. With Haiku 5.5 on the way in the coming weeks, the family is complete for every budget and speed. Altman and Musk backed the call; Zuckerberg dismissed it. That the warning also justifies entrenching the advantage is why it costs so little to sign it: no lab truly slows down when slowing down hands the release cycle to the rival.

## The paradox of power

The prospectus hides a contradiction that deserves slow reading. To make AI safe, Anthropic says, control and capital must be concentrated in ever fewer hands; in fact, Anthropic says there are no better hands than its own. Ironic, to ask for the monopoly of prudence while denouncing the monopoly of power: welcome to the logic of frontier labs, where every argument serves twice, and the plot twist always benefits them.

Dependence on others' computing adds another crack to the story. If access to third-party computing is cut, becomes more expensive or is severed, the document warns, the business suffers. Translated: the company that wants to save humanity from AI depends on clouds it does not control to train the AI it must save us from. The paradox is not an oversight in the prospectus, it is the exact photograph of the sector.

Basically, Anthropic tells you to your face: **I am the devil, I am creating a weapon to destroy humanity, but I am a nice guy, and I want to save you, if you give me all the power and a modest monthly fee.** It plays thug and cop, but it also wants the power of judge, jury and executioner.

## Listing the apocalypse

Asking for billions while warning of the end of the world is a move that only works once, and Anthropic is playing it now. The message to investors is twofold: if AI changes everything, whoever leads it captures the biggest prize in economic history; if it goes wrong, at least they warned you on page 80. With the debut expected after the midterms and SpaceX as the yardstick at $1.77 trillion, the market will price that ambivalence with real money. The prospectus devotes twice as many pages to risks as to the business, and still asks for the highest valuation in stock-market history. That disproportion is the real financial headline: never was so much paid for a company that warns so much. Existential risk and existential opportunity in the same prospectus, signed by the same hands.

We will have to wait for the debut to know how the market prices that ambivalence. If the price withstands the scare, a new manual for listing dangerous technology will be born. What no one disputes anymore is that the industry has changed phase: when the most cautious lab devotes twice as many pages to risks as to the business, safety has stopped being a department and become the balance sheet. And balance sheets, as is well known, are also audited, and this one promises to be the most scrutinized of the decade.

**[The Hugging Face hack and the powerful message it sends the world](https://observatorioblockchain.com/en/cybersecurity/hugging-face-hack-mensaje-mundo-empresarial/)**
