Claude Mythos Preview: Anthropic Just Chose Not to Sell Its Strongest Model

On April 7, 2026, Anthropic’s Project Glasswing announcement introduced something strange: they had trained a model stronger than Opus 4.6 and GPT-5.4, called Claude Mythos Preview, and they were not going to make it generally available.

The standard frontier AI template goes: announcement, benchmark screenshots, API opens, pricing posted, valuation bumps. That’s not what happened.

Anthropic’s internal red-team evaluation pointed in the other direction: Mythos is strong enough that Anthropic concluded it shouldn’t be on the open market. How strong, specifically?

• SWE-bench Verified: 93.9%
• SWE-bench Pro: 77.8%
• GPQA Diamond: 94.6%
• Thousands of zero-day vulnerabilities across every major OS and browser
• One OpenBSD bug present for 27 years that Mythos is the first tool to find

Those scores are a double-digit lead over Opus 4.6 and GPT-5.4. And Anthropic said no.

What Project Glasswing Actually Is

Project Glasswing is the structure Anthropic built around Mythos. The logic is simple: give a small set of organizations responsible for maintaining critical infrastructure early access, so they can close vulnerabilities before attackers get comparable capabilities.

The publicly confirmed launch partners include:

• AWS, Apple, Google, Microsoft, NVIDIA
• Broadcom, Cisco, Palo Alto Networks, CrowdStrike
• JPMorgan Chase
• The Linux Foundation
• Anthropic itself

Beyond that, around 40 more organizations get access at Anthropic’s discretion. The criterion is whether they build or maintain critical software infrastructure. Fortune reported pricing at $25/M input tokens and $125/M output, several times Opus 4.6’s standard rate. The pricing is not really the story. The list is.

Why This Actually Matters

For the past two years, the favorite hypothetical in AI safety circles has been: “what happens when a model can autonomously find and weaponize high-quality zero-days?” The answer has stayed vague. Mythos moved that timeline forward.

But the model capability isn’t the most interesting part. The interesting part is that Anthropic made a deliberate commercial choice not to sell it.

Three layers to pull apart.

One: this is a precedent. Frontier model economics has run on a simple rule: stronger model, faster to market, higher valuation. Anthropic just inverted it. Bruce Schneier wrote the cleanest framing of this: “Anthropic is effectively admitting they built something they shouldn’t sell as a product.”

Two: this is the first real test of frontier safety policies. Simon Willison’s take is that various responsible scaling commitments AI labs have signed have never been stress-tested in a “model is trained, pricing could ship tomorrow” situation. Mythos is the first. The meaning isn’t just what happens to this model. It’s that future versions of the same decision now have a reference point.

Three: this is a commercial risk. Anthropic just pulled their most commercially valuable asset off the shelf. Shareholders have opinions. Competitors respond. Internal teams get restless. We’ll see that response below. OpenAI’s GPT-5.4-Cyber landed within days.

How Mythos Found a 27-Year-Old OpenBSD Bug

This detail deserves scrutiny because it’s what separates marketing from reality.

Both Fortune and Ctech make the same point. Mythos doesn’t just find vulnerabilities; it writes working exploits.

The old AI-assisted security research workflow was:

1. AI reads code
2. AI flags suspicious regions
3. Humans write the exploit to verify

Mythos eats step 3 too. It runs the full loop autonomously, from static analysis to exploit to copying a secret out of memory into another directory. Anthropic’s internal tests handed it 50 crash categories and asked it to develop working exploits. It worked through them.

If you’re wondering whether this is marketing, I’d say probably not. The launch partner list includes Apple, Google, Microsoft, and the Linux Foundation. Those are organizations with the strongest security teams in existence. They wouldn’t waste time on the partnership if Mythos couldn’t do what it claims.

Open Questions

Things I haven’t seen answered yet but that the community should keep tracking:

• What’s the relationship between Mythos Preview and the real Mythos? Preview suggests a full version is coming. Will it be stronger, or delivered as API-only?
• Will Glasswing expand the list? Forty organizations sounds like a lot until you compare it to the global surface area of software that needs protection
• How long until an open-source equivalent appears? Zhipu’s GLM-5.1 shipped in the same week and scored 58.4 on SWE-bench Pro. If that trend continues, Mythos-class capability will eventually exist under an MIT license, and “not selling it” will no longer be Anthropic’s call to make

That last point is actually the main one. Anthropic can choose not to sell Mythos because they’re currently the only lab that can train it. How long does that position hold? A few months? Years? Forever? The answer shapes how frontier-AI safety policy writes itself for the rest of the decade.

What I’m more confident about: starting this year, “what model did you train” and “how did you choose to deploy it” become two separate questions. Mythos draws that line clearly.