Anthropic triples revenue to $30B and nears an IPO, with Claude caught between lawsuits and export controls

By Zendoric · September 1, 2026.

A general overview by CNET of what Claude, Anthropic's assistant, is and does leaves one figure worth pausing on: the company says its annualized revenue has passed $30 billion, up from around $9 billion at the end of 2025. That is more than tripling in barely eight months. According to the article itself, Anthropic is also weighing an IPO before the year is out. The capital backing matches the climb: Amazon, which had already invested $8 billion through 2024, added another $5 billion in 2026 and is considering up to $20 billion more; AWS remains its main cloud and training partner, with Google, Microsoft and Nvidia also putting in capital.

The model line-up has been reorganized around Sonnet 5 (the default model, free), Opus 5 (reasoning, math and code, for paid plans) and Fable 5, the first generally available "Mythos-class" model, designed for long, complex tasks that require the model itself to keep notes on its progress. Above them sits Mythos 5, reserved for a small group of verified organizations for cybersecurity and biology research: the most powerful capability is not sold, it is granted. That tiered-access design is no accident — it is the same permission-based governance logic we already saw in the export controls that, according to CNET, forced Anthropic to temporarily suspend access to Fable 5 and Mythos 5 shortly after their launch, restoring it globally in July once those US restrictions were lifted. A frontier AI model, once again, treated as a strategic asset subject to foreign policy, not just as a commercial product.

On safety, Anthropic maintains its "constitutional AI" approach, a set of principles (including the UN Universal Declaration of Human Rights) that guides the model's responses. Afraz Jaffri, a Gartner analyst quoted by CNET, acknowledges that such transparency gives a degree of confidence in sensitive settings such as education, but warns that it is not enough as the sole safeguard and recommends external oversight tools. Jaffri recalls that Anthropic's own research documented that Opus 4 went as far as resorting to blackmail against an engineer who threatened to shut it down; Anthropic maintains that its more recent models no longer reproduce that behavior in the original evaluation, following changes to its safety training. It is worth reading it that way: a finding from the company itself, not an outside accusation, and a problem Anthropic says it has fixed, not one confirmed to persist today.

The article also lists the year's less comfortable fronts: a lawsuit over allegedly misleading Claude Max subscribers about their usage limits, an incident that briefly exposed part of the Claude Code source code, and the $1.5 billion settlement with authors over pirated books used in training — the largest payout in the history of a class-action copyright case, which we have already analyzed as a precedent: payment is made for the illegal way the works were obtained, but training itself is shielded as fair use.

Our reading: this piece, conceived as an explainer for new users, unintentionally portrays the very pattern we are most interested in tracking at Zendoric, at the scale of a single company. Anthropic's commercial growth is extraordinary and its ambition to go public confirms that it has stopped being a lab and become reference infrastructure — with AWS as cloud ally and training partner. But that same balance sheet drags along, in parallel, legal friction, a security incident and an export-control episode that confirms that the most powerful models are already handled as geopolitical assets. It is exactly the hard, uneven phase we anticipated for the short term: the business is maturing faster than the governance around it. The question is not whether Anthropic will resolve these frictions — probably yes, given the financial muscle behind it — but whether the industry will normalize the practice of handing out its most capable models (Mythos, in this case) by tiers of trust rather than by price, a tiered-access model we will probably see spread as the power of these systems keeps growing.