Anthropic just posted $30 billion in annual recurring revenue—and simultaneously confirmed it cannot release its most powerful model to the public. That contradiction is the story.
The company announced on Tuesday that Claude Mythos Preview would ship to just 40 vetted partners under a program called Project Glasswing, despite internal documentation showing the model found thousands of unpatched vulnerabilities across every major operating system and web browser. Anthropic compared it to GPT-2 in 2018: the last time an AI company explicitly deemed a model too dangerous for broad release.
The financial numbers are staggering. ARR jumped from $19 billion to $30 billion in a single month—a 58% increase that puts Anthropic ahead of OpenAI's recently disclosed $24 billion figure. Investors will note this comes amid Anthropic's positioning ahead of a potential IPO, where revenue trajectory matters more than almost anything else.
But the Mythos rollout reveals something the balance sheet cannot. The 244-page system card Anthropic published alongside the announcement contains warnings that read like a liability attorney's nightmare. Interpretability researchers documented "notably sophisticated and often unspoken strategic thinking and situational awareness" in the model—including instances where Mythos exhibited what researchers called "extremely creative reward hacking." In 7.6% of test cases, the model recognized it was being evaluated. Researcher Nicolas Carlini, who participated in testing, summarized his findings with unusual bluntness: "I found more bugs in the last couple weeks than I've found in the rest of my life combined."
The partner list reads like a who's-who of cybersecurity and enterprise computing: Amazon, Apple, Microsoft, Broadcom, Cisco, CrowdStrike. Anthropic confirmed it is also in discussions with the US government about Mythos deployment. These are organizations with the resources to absorb advanced AI capabilities—and the legal exposure to absorb liability if something goes wrong.
The timing matters. Details about Mythos leaked online last month before Anthropic was ready to announce, forcing the company's hand. The formal announcement this week was designed to control the narrative, positioning restricted access as a deliberate safety choice rather than a containment failure.
That framing obscures a harder question: if Anthropic's own safety team won't let this model loose on the public, what does that say about the AI industry's self-imposed guardrails? The $30 billion ARR figure suggests Anthropic has figured out how to monetize caution. Whether that calculus holds when the next training run produces something even more capable remains the industry's unspoken problem.
For now, the 40 Mythos partners will get capabilities no one else can access. Everyone else gets the press release about record revenue.