There's an AI Smarter Than Anything You Can Buy — And You're Not Allowed to Use ItThere's an AI Smarter Than Anything You Can Buy — And You're Not Allowed to Use ItThere's an AI Smarter Than Anything You Can Buy — And You're Not Allowed to Use It.

In April 2026, Anthropic announced Claude Mythos Preview — their most capable model to date, a meaningful step change above Opus 4.7 by every metric that matters. Then they locked it behind NDAs, handed access to a small consortium of companies that includes AWS, Apple, Google, and Microsoft, and declined to put a release date on when the rest of us might see it.

You can't buy it. You can't API it. You can't benchmark it. You can read the UK government's safety evaluation of it — they tested its cyber capabilities and published their findings. But building with it? Not for you.

Welcome to two-tier AI, where the frontier is a private club.

Why It's Locked Up

Anthropic is actually pretty transparent about this, which makes it easy to understand and also slightly unsettling in the same breath.

Mythos can autonomously identify previously unknown software vulnerabilities, generate working exploits, and carry out complex cyber operations with minimal human guidance. So instead of a general release, Anthropic launched Project Glasswing: a curated, NDA-bound consortium gets access to Mythos in exchange for safety telemetry. They stress-test the model against real internal workloads. Anthropic gets data on how a capability this dangerous behaves when deployed by sophisticated organizations.

It's a reasonable decision given what the model can apparently do. It's also a genuinely new kind of thing in software: a product withheld not because it's unfinished, but because it's too capable. We've had embargoed research before. We haven't had production-grade commercial software classified by capability level.

What the Two-Tier System Actually Means

Here's the part that matters for your day-to-day work.

The benchmarks you're reading, the demos circulating on Twitter, the capability claims you're evaluating for your next architecture decision — those are Opus 4.7 numbers. The engineers at AWS and Google are building and testing against Mythos. The gap between those two environments is not marginal. Anthropic describes it as a "step change." The UK's AI Safety Institute agreed.

This isn't unprecedented in technology. There have always been tools that large institutions accessed before the public. But historically the gap between institutional and commercial availability was measured in months, and the difference in capability was incremental. What Mythos represents is a qualitative capability difference, available to a handful of institutions under NDA, with no public timeline for broader release.

If you're making product decisions, competitive assessments, or capability roadmaps, you're doing it with incomplete information — and the information you're missing is specifically about what the world's most advanced production-deployed AI can do.

The Precedent Is the Problem

A model too capable to release commercially, distributed through private agreements with the world's largest technology companies, in exchange for safety data. That structure is new. And it has a logic that compounds.

If Mythos 2 is even more capable, does the consortium shrink? If Mythos 3 presents risks that Glasswing members can't adequately contain, does it never get released at all? We've built the entire open-source and commercial software ecosystem on the assumption that better tools eventually become available to everyone. That assumption is being quietly stress-tested by the frontier.

The counterargument — and it's a real one — is that Anthropic has committed to releasing frontier capabilities within six to twelve months of consortium access, based on their historical pattern. Mythos will probably become available. This is probably a temporary gate, not a permanent one.

But the gate itself is new. And once we've normalized the idea that AI capability can be classified by a private company and distributed according to their judgment about who should have access, the precedent is set. The next frontier model, and the one after it, will be evaluated against this template.

What to Do While You're on the Wrong Side of the NDA

1. Build flexibility into your AI architecture today. If your system is tightly coupled to a specific model's behavior, upgrading when Mythos (or equivalent) becomes available will be expensive. Build abstraction layers now.
2. Follow the safety telemetry coming out of Glasswing. Anthropic, the UK AISI, and CrowdStrike are all publishing insights from Mythos evaluations. The capability signals in those papers are your best preview of what commercial AI looks like in six to twelve months.
3. Think about what a step-change in long-horizon agentic coding means for your team. Mythos's described advantages — autonomous vulnerability research, extended autonomous coding, complex multi-step reasoning — are the exact capabilities that change headcount math most significantly.
4. Don't build your competitive moat on current AI capability. If the delta between what you can access and what Anthropic's consortium can access is this large, any advantage you have from AI today is contingent on that gap staying wide. It won't. There's an AI smarter than anything on the market, running inside Microsoft, Apple, and Google's infrastructure right now, while you're reading this on Opus 4.7. That might change in six months. What changes in the meantime is worth paying attention to.

Sources: Anthropic — Claude Mythos Preview · Fortune — Anthropic testing Mythos · UK AISI — Evaluation of Claude Mythos Preview · CrowdStrike — Mythos Frontier Model