Goldman Sachs and Morgan Stanley have started internal trials of Claude Mythos Preview, the most powerful model Anthropic has built. They've kept it from the public because their own internal checks found it was simply too capable to release broadly. These trials, which Bloomberg and Reuters confirmed on April 10, follow a strange sequence of events: a pre-launch briefing between the White House and several tech CEOs, an emergency Treasury meeting, and the discovery of a 27-year-old software vulnerability that had escaped every automated tool before it.
A Model for the Gaps
Anthropic introduced Claude Mythos Preview as part of "Project Glasswing," an initiative aimed at defensive cybersecurity. It isn't a specialized security tool, but a general-purpose model that showed dangerous depth in coding and reasoning during its safety evaluations. Mythos identified thousands of severe vulnerabilities across major browsers and operating systems. One was a remote-crash flaw in OpenBSD that had lasted 27 years and survived millions of scans; another was a 16-year-old bug in FFmpeg that five million automated attempts had missed. The model can even turn these flaws into working exploits on its own, a feat Anthropic says its previous flagship, Opus 4.6, almost never managed. For now, access is limited to about 40 groups, including Apple, AWS, Google, Microsoft, and JPMorgan Chase—firms that sit inside the world's critical infrastructure. Preview pricing stands at $25 per million input tokens and $125 per million output.
Regulators Signal Systemic Risk
Before launch, Vice President JD Vance and Treasury Secretary Scott Bessent called CEOs from Anthropic, Google, Microsoft, and OpenAI to discuss AI security and how to defend against automated cyberattacks. Then, on April 8, Bessent and Federal Reserve Chair Jerome Powell held an urgent session at the Treasury with the heads of Goldman Sachs, Morgan Stanley, Bank of America, Citigroup, and Wells Fargo. The takeaway was that Mythos-class models can find vulnerabilities in hours rather than months—a pace that's faster than any major bank can patch. JPMorgan’s CEO couldn't make the meeting, but his recent shareholder letter called cybersecurity a grave risk and warned that AI will likely make the problem worse.
The Performance Tax on Consumers
Since late February, weeks before any mention of Mythos, heavy users of Claude.ai and Claude Code started noticing slower responses and shallower reasoning. They also found themselves hitting usage limits much faster than before. The complaints got louder after Project Glasswing rolled out on April 7. Analysis on developer forums points to a 67% drop in "thinking depth" metrics. Anthropic hasn't confirmed if they've diverted GPU resources to the new model, pointing instead to routing bugs or changes in default effort settings. But for sophisticated users watching the data, the timing makes those explanations hard to swallow.
The Real Investment Signal
The story of Mythos is where the bullish consensus gets sloppy and the real questions for investors actually begin.
Mythos is handling four jobs for Anthropic at once. It shifts the company from being a model vendor to a critical infrastructure partner. It deepens their hold on the coding and agentic markets. It creates a "responsible" brand that sells better than openness in regulated industries. And it acts as a fear-based demand accelerator that pulls enterprise budgets forward. The reaction from the Fed and Treasury confirms this: no bank that matters to the system can afford to sit still when regulators are warning about machine-speed exploit discovery.
Still, the case for the bulls has clear limits. Anthropic’s own technical reports say the cyber leap came from general improvements in reasoning, not specialized training. This suggests competitors will hit the same marks before long. Mythos is also served through Vertex AI, Bedrock, and Microsoft Foundry, which means the cloud giants get a cut every time Anthropic wins. While a multi-gigawatt compute deal with Google and Broadcom shows strength, it also shows that Anthropic’s moat is built partly on contracts rather than unique technology. Furthermore, Bloomberg reports the Pentagon labeled Anthropic a supply-chain risk—a move the company is fighting in court—even as it tries to position itself as the government’s trusted cyber partner. The "responsible national champion" image is a powerful selling point, but the ground it sits on is contested.
Anthropic’s revenue run-rate has passed $30 billion, up from $9 billion at the end of 2025. The number of customers spending over $1 million a year doubled to more than 1,000 in just two months. It’s a real enterprise franchise forming at high speed.
Anthropic isn't winning "AI" in a general sense. It is winning the process of turning frontier model capability into a sense of premium urgency for large companies. The claim is narrower, but it is much stronger. Mythos is the sharpest version of that strategy yet. Whether this becomes a durable moat or just a well-timed narrative advantage is the only question that matters for investors.
not investment advice
Sources: Anthropic Official Announcement: Project Glasswing and Claude Mythos Preview — https://www.anthropic.com/glasswing
Anthropic Project Glasswing page (partners, commitments) https://www.anthropic.com/project/glasswing