Why Anthropic's Most Powerful AI Is Only Available to 50 Companies

Anthropic has built a model it considers too dangerous to release. Claude Mythos Preview was launched on April 7, 2026 and immediately went behind a closed door. No public API. No pricing page. No timeline for general availability. Instead, roughly 50 organizations were hand-selected to access it, all of them chosen because they defend the kind of infrastructure billions of people depend on.

The selectivity is not marketing. It is a direct response to what Anthropic found during testing.

What Makes Mythos Different

Claude Mythos Preview is not just a stronger version of Claude Opus. It represents what Anthropic describes as a "step change" — a model that performs in a category above its predecessors specifically when it comes to finding and exploiting software vulnerabilities. According to Anthropic's own red team writeup, Mythos is capable of identifying zero-day vulnerabilities in every major operating system and every major web browser. It reported 271 zero-day vulnerabilities in Firefox 147 alone. Some of the bugs it surfaced were decades old and had gone undetected by human researchers.

On standard AI benchmarks, the numbers are similarly striking: 97.6% on USAMO 2026 mathematics, and roughly 93.9% on SWE-bench Verified, which tests a model's ability to resolve real software engineering issues. But it is the cybersecurity capability that drove Anthropic's access decision. A model that can autonomously discover and build working exploits from vulnerabilities is what security professionals call dual-use — equally useful for defense and for attack.

Project Glasswing and the 50-Company List

Anthropic's response was Project Glasswing, a structured program that channels Mythos access exclusively to organizations already responsible for critical infrastructure. The twelve launch partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Beyond those, over 40 additional organizations were invited, all chosen for building or maintaining software that underpins shared systems at scale.

Anthropic committed $100 million in usage credits to cover the program and donated $4 million directly to open-source security groups. The logic is straightforward: give defenders access to the model's capabilities before attackers develop equivalent ones independently, and use the window to patch as many vulnerabilities as possible. The theory is sound. The execution hit an early obstacle; Bloomberg and TechCrunch reported that unauthorized access was obtained through a third-party vendor environment within roughly two weeks of launch.

What This Means for Everyone Else

For most businesses and everyday users, Mythos is not a product you will use anytime soon. Anthropic has been explicit that it has no plans to make it generally available. Claude Opus 4.7, released in mid-April, remains the highest-capability model accessible through standard API and consumer subscriptions and is genuinely powerful for the vast majority of use cases.

The deeper implication is about the direction of AI development. For the past several years, the standard arc for a new frontier model was: announcement, benchmarks, public API within weeks. Mythos broke that arc. It is the first model in the modern AI era where the headline capability numbers belong to something the public cannot access. If the pattern holds, future models with similarly sensitive capabilities may arrive the same way: restricted first, with general availability as a secondary consideration, or possibly never.

That is a meaningful shift in how power over frontier AI is distributed. The 50 companies with Mythos access include some of the largest technology and financial institutions in the world. The gap between what they can do with AI and what independent developers or small businesses can do just widened — not because of price, but because of a deliberate policy decision about who should hold certain capabilities and when.

Whether Anthropic's bet pays off depends on whether the defensive use of Mythos can outpace the offensive capabilities it is helping to develop across the broader AI ecosystem. The early leak suggests the window is narrower than anyone hoped.

This analysis is based on reporting from DEV Community, Anthropic, and TechCrunch.

Image courtesy of FlyD and Unsplash.

This article was generated with AI assistance and reviewed for accuracy and quality.