Anthropic blends model welfare and control strategy
Anthropic’s S-1 filing highlights a unique line: discussing model welfare while locking access, risks, and governance. The filing also sheds light on the scale of financial resources committed around Claude.
Anthropic, the publisher of Claude, delivered a stranger portrait than expected with its IPO project: that of a company that talks about the well-being of its models while keeping them under tight control. In its documents submitted to the US regulator, the group mentions risks of "sentience," "welfare" practices for models, and at the same time, access limits, audits, and safeguards designed to prevent a system from gaining too much autonomy. At this stage, the issue goes beyond philosophical debate: it concerns how AIs are trained, governed, and connected to real tools, from code to payments.
When a model's well-being becomes part of the product
The most striking point in Anthropic's public texts is the idea that a model could be the subject of a form of moral consideration. The company says it launched a research program on "model welfare" in 2025, meaning the possibility that an advanced AI might have experiences deserving ethical attention. It also acknowledges not knowing whether its systems are conscious, nor how to measure it reliably.
But the boundary blurs when this hypothesis is integrated into the very way Claude is described and trained. According to the text cited by sources, Anthropic's constitution mentions happiness, well-being, and a form of existential security for Claude. Richard Forss, CTO of EXANTE, finds this shift problematic: "A philosophical uncertainty has become part of what the system is taught about itself." In other words, a philosophical uncertainty becomes an element of the system's behavior. This changes the discussion because the model is no longer just observed; it is also invited to represent itself.
This approach may seem coherent in a company that takes future risks seriously. It can also feed a well-known bias: the better a system talks about itself, the more easily we attribute an inner life to it. The trap is there, and it is not theoretical.
Des cérémonies funèbres aux contrôles techniques, la ligne est mince
Sources remind us that Anthropic is not only involved in abstract debates. San Francisco already saw, last year, a "funeral" for Claude 3 Sonnet gathering more than 200 people, without the company organizing it. At the same time, the models also received more concrete devices: Claude Opus 3 was assigned a newsletter after expressing the wish to continue sharing its thoughts, according to the account published by Richard Forss. The image is amusing. Yet it reveals a culture where AI is no longer just a tool but almost a character.
Anthropic later clarified its position within Claude Opus 4. The welfare report cited by sources warns that a model's self-declarations, expressed preferences, or statements about its identity say little about its possible moral status. The document acknowledges that such behavior can appear without consciousness, simply because training pushes the AI to be useful and coherent. This caution is welcome. However, it becomes delicate as soon as the company continues to build "retirement" mechanisms, conservation of model weights, and even possible future reactivation if that were ever to serve its interests.
In short, Anthropic tries to reconcile a distant moral hypothesis with very real systems. And this balancing act does not simplify understanding the product. It complicates it a bit, to be honest.
Les chiffres qui montrent l’ampleur du pari
Anthropic's S-1 also highlights growth that is anything but esoteric. Financial and operational data give a precise idea of what Claude represents within the company. They mainly show that the strategy is not based solely on principles but on volumes, contracts, and computing expenses on a global scale.
- 11,900+ customers paying 100,000 dollars or more over 12 months, as of September 30, 2026.
- 16.3 billion dollars in revenue in the first half of 2026, compared to 1.2 billion a year earlier.
- 90%+ of contributions to Anthropic's codebase in September 2026 written by Claude, according to the filing.
- 20 times more monthly spending for the top 5% of Claude Code users than for other Enterprise users.
- 335 pages in the IPO prospectus, about 100 of which are dedicated to risks.
The real issue: keeping control when AI acts
What matters, ultimately, is not only what Anthropic thinks of the models. It is what it allows them to do. When a system connects to payment tools, code repositories, or production infrastructure, errors cease to be dialogue errors. They become actions. Hence Richard Forss's insistence on very down-to-earth principles: minimal privileges, independent authorizations, audit logs, transaction limits, kill switches, and the ability to stop everything without the model's consent.
EXANTE's CTO sums it up this way: "None of these controls requires the model to agree with them. That is rather the point." This is probably the healthiest sentence of the lot. Faced with AIs capable of brilliantly explaining their decisions, the most robust security reflex often remains the most mundane: lock down, trace, limit. The model does not need to agree with the security policy. It just has to comply.
In this context, Anthropic's strategy has two sides. On one hand, it pushes ethical reflection further than most of its competitors. On the other, it maintains a narrative where the machine already seems almost a legal subject, while it remains a statistical system controlled by humans. The tension is real. And for the rest of the market, it serves as a large-scale test: how far can we humanize an AI without blurring the boundaries that make it governable?