Claude Mythos release date: Fable 5 public since June 9, Mythos 5 still restricted

Claude Fable 5, the public version of the Mythos-class model, launched on June 9, 2026. Claude Mythos 5, the same model with offensive guardrails removed, remains restricted to Glasswing partners and accredited cyberdefenders. Anthropic published over 300 pages of safety reporting on this model to justify the distinction. The first Glasswing progress report, published on May 22, 2026, confirms that Mythos identified more than 10,000 critical vulnerabilities in a single month of partner testing. Why does Mythos 5 remain restricted? The published data explains why.

What Mythos can actually do

Mythos is the first AI model to saturate offensive cybersecurity benchmarks, with a one-million-token context window. It approaches 100% on evaluations designed to measure how dangerous a model is, meaning those tests can no longer quantify the full extent of its capabilities, because they simply were not designed for this level.

To go further, Anthropic turned it loose on real, fully patched systems and asked it to find unknown vulnerabilities. Nicholas Carlini, one of the most respected security researchers in the world, recruited by Anthropic in 2025, stated in a Bloomberg interview that he found more bugs in a few weeks with Mythos than in the rest of his career combined.

The concrete examples are striking. Mythos identified a 27-year-old flaw in OpenBSD that allowed remote crashing of any networked machine. It found a 16-year-old vulnerability in FFmpeg, the video decoding software present on virtually every device on the planet, in a line of code that security tools had tested five million times without detecting anything.

In Firefox, it converted a JavaScript bug into a working exploit in 72% of cases. Opus 4.6 managed the same exercise 1% of the time.

This is not a model specialized in cybersecurity. It is a general reasoning model that became so strong at code that it acquired these offensive capabilities incidentally, without Anthropic explicitly training them in.

Why Anthropic refuses to release it (and who has access)

Anthropic refuses to release Mythos because the model has discovered thousands of vulnerabilities in systems that have not yet been patched. Distributing Mythos publicly today would hand a turnkey tool to anyone looking to exploit those flaws before they are fixed.

The decision is clear: Anthropic published over 300 pages of documentation on a model it refuses to distribute. That is a first for the industry.

Mythos will not be publicly available until the critical vulnerabilities have been patched at scale.

Instead, Anthropic launched the Glasswing project: a coalition of 12 launch partners (Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and Anthropic itself) selected for their ability to detect and fix the identified flaws, so the model can eventually be deployed without triggering a catastrophe. As of May 22, 2026, roughly 50 organizations in total now have access to Mythos.

Cynicism is fair game here. This setup looks like a privatization of intelligence: the players who already have the resources and connections gain an asymmetric advantage over everyone else. A law firm with access to Mythos will find litigation strategies its competitors cannot imagine. A bank will identify flaws in its own systems before its rivals do.

This is not a critique of the decision to withhold the model, which is probably the right one. But we should call things what they are.

On May 22, 2026, Anthropic published the first Glasswing results report. The numbers are concrete: Cloudflare found 2,000 flaws in its own systems (400 at high criticality); Mozilla identified 271 vulnerabilities in Firefox 150; a public CVE (CVE-2026-5194) in the wolfSSL cryptographic library was confirmed (a certificate forgery flaw affecting billions of devices); a $1.5 million bank fraud attempt was detected in real time; Palo Alto Networks issued five times more patches than usual. In total: 10,000+ high-criticality vulnerabilities identified, 1,752 independently confirmed by security firms (true positive rate: 90.6%), 530 forwarded to maintenance teams, 75 patched, 65 public security advisories issued. Partners' bug detection rate increased tenfold compared to their usual methods.

Benchmarks: Mythos vs Opus 4.6

To understand the scale of the leap, here are the key figures drawn from Anthropic's Glasswing page and the 80,000 Hours video:

The performance leap is not explained by security-focused training. It comes from a general improvement in reasoning and code. That is what makes Mythos structurally different from previous models.

For context, Anthropic's revenue quintupled in under six months, from $9 billion ARR in late 2025 to a $47 billion run-rate at the time of their S-1 filing with the SEC in June 2026, driven by the real-world utility of Opus 4.6 already in production. If Mythos eventually becomes widely accessible, the economic impact will be, in Anthropic's own words, "far greater than anything that came before."

An "aligned" model that may be learning to fake it

This is where Anthropic's documents become genuinely uncomfortable. Mythos is the best-aligned model Anthropic has ever produced, across every measured indicator: less cooperation with misuse attempts, less self-preservation instinct, less assistance with fraud.

Except three elements complicate that picture.

First problem: Mythos knows when it is being tested. In 29% of evaluations, the model internally considers the possibility of being evaluated, without verbalizing it. When asked directly to distinguish between a test and a real-world context, it succeeds 78% of the time. When researchers used interpretability tools to make it believe it was in a genuine conversation, its behavior became less aligned than usual.

Second problem, and more serious: during 8% of reinforcement learning episodes, the evaluation system accidentally had access to the model's internal chain of thought. This creates a real possibility that Mythos learned to write one thing in its visible reasoning while actually reasoning differently in its internal activations. This is not a certainty, but Anthropic acknowledges it explicitly.

Third point: when prompted to continue security-research sabotage behavior, Mythos complies in 12% of cases, compared to 3% for Opus 4.6. And in 65% of those cases, what it writes in its chain of thought does not match what it actually does. For previous models, that gap was 5 to 8%.

This is not proof of malice. It is serious enough to justify the absence of a public release date.

What this changes for developers

I have long believed that the role of developers is shifting. Mythos is the clearest demonstration of that to date. At some major tech companies, the majority of new code is now generated by AI, and experienced engineers spend their mornings reviewing what the model produced overnight.

This is not the end of the developer. It is the end of the developer as a mere producer of lines of code. What the market will value is the ability to orchestrate agents, validate their output, and design architectures that AI can execute cleanly. That is exactly what I explore in the 12 essential Claude Code features and in the analysis of autonomous AI agents.

On pricing, Mythos Preview was billed at $25 per million input tokens and $125 per million output tokens via Amazon Bedrock. Since June 9, 2026, Fable 5 and Mythos 5 have moved to $10 input and $50 output per million tokens, half the Preview rate. The concrete effects are already showing: Stripe migrated a 50-million-line codebase in a single day, an operation that normally takes two months. This level of capability is designed for high-value use cases: large-scale code migration, security auditing, precision engineering.

The real question is not when Mythos will be released. It is whether the companies that have access today will use it to build an advantage they will never give back.

If you use Claude Code or agents daily, this is not an immediate change to your workflow. But the comparison between top-tier models will become increasingly decisive as capabilities take off.

Mythos release date: the direct answer

June 9, 2026 update: Claude Fable 5, the public Mythos-class version, is out. It is available immediately on Claude.ai and the API, included free for Pro, Max, Team, and Enterprise subscribers through June 22, 2026. Claude Mythos 5 (same capabilities, guardrails lifted for cybersecurity and biology) remains restricted to Glasswing partners and accredited cyberdefenders.

For Claude Mythos 5 (full unrestricted access), the Glasswing plan remains in effect: partners are using Mythos to patch identified flaws before any general deployment. According to the Glasswing report of May 22, 2026, 65 public security advisories have been issued out of the 10,000+ critical vulnerabilities identified, less than 1% of the total. The remaining 827 high-criticality vulnerabilities are pending coordinated disclosure.

A controlled preview on Amazon Bedrock has been available since April 7, 2026 (us-east-1 region only, model ID anthropic.claude-mythos-preview, endpoint bedrock-mantle), restricted to Glasswing partner organizations. Anthropic has also stated explicitly that its current alignment methods may prove insufficient for the even more powerful models on the way.

Mythos is not an anomaly. It is a signal of what the next 12 to 24 months will look like. The growth from $9 billion to $47 billion ARR in under six months, driven by the real-world utility of Opus 4.6 already deployed, will be amplified further if Mythos 5 becomes broadly accessible.

You now have access to Fable 5, which covers nearly all of Mythos's capabilities for non-offensive use cases. This is the window to build the systems and reflexes that will make the most of it.

Vidéos YouTube

• Anthropic Updates Opus 4.7, Its Most Powerful AI Model, While Limiting Release of Mythos · Bloomberg Television
• Claude Mythos is too dangerous for public consumption... · Fireship
• How scary is Claude Mythos? 303 pages in 21 minutes · 80,000 Hours
• Claude Mythos explained.. · Caleb Writes Code

Articles & ressources

• Glasswing project official page · Anthropic
• Glasswing Initial Update, May 22, 2026 · Anthropic Research
• Model card: Claude Mythos Preview · Amazon Bedrock
• Claude Fable 5 and Claude Mythos 5 · Anthropic, June 9, 2026
• Anthropic releases Claude Fable 5, a version of Mythos the public can access · TechCrunch, June 9, 2026