Claude Mythos 5: What It Is, Access, and How It Compares

Last updated: July 25, 2026

Claude Mythos 5 is the most capable — and most restricted — model in Anthropic’s Claude family. Released alongside Claude Fable 5 on June 9, 2026, Mythos 5 shares the same underlying model family but removes some of the safety classifiers that Fable 5 applies to high-risk domains. The result: the highest raw capability tier Anthropic has shipped, available only to vetted partners through Project Glasswing and trusted-access programs.

Quick answer: Claude Mythos 5 is a restricted-access Mythos-class model. It is not available through the public API, Claude.ai, or any self-serve channel. On standard coding and knowledge-work benchmarks, it shares the same scores as Fable 5 because they use the same underlying model. Where Mythos 5 diverges is in safeguarded domains — cybersecurity, biology, and chemistry — where Fable 5’s classifiers restrict or reroute requests. For most users, Fable 5 or Opus 5 delivers the same core intelligence with fewer barriers.

What is Claude Mythos 5?

Claude Mythos 5 is Anthropic’s highest-capability frontier model, released on June 9, 2026 as part of the same launch as Claude Fable 5. The two models share the same underlying model family — the difference is not raw intelligence but access controls and safety classifiers.

Think of it this way: Mythos 5 is the unrestricted research-grade version. Fable 5 is the production-grade version with guardrails added for domains where advanced AI capability creates real-world misuse risk.

Anthropic’s positioning is deliberate: Mythos 5 exists because the model’s capabilities in cybersecurity, biology, and chemistry are strong enough to create significant dual-use risk. Rather than weakening the model to make it safe for general release, Anthropic ships the same capability broadly as Fable 5 (with classifiers) and reserves the unrestricted version for vetted partners who need it for defensive and research purposes.

Mythos 5 vs Fable 5: what is actually different?

This is the most common question, and the answer is more nuanced than “Mythos 5 is smarter.”

The critical insight: on normal coding, knowledge work, reasoning, and vision tasks, Mythos 5 and Fable 5 produce the same results. Anthropic’s published benchmark table uses a combined “Fable 5 / Mythos 5” column because the scores are identical for non-safeguarded evaluations. The divergence appears only when a request touches cybersecurity, biology, chemistry, or distillation domains — areas where Fable 5’s classifiers may refuse, reroute to Opus 4.8, or return a summarized response instead of the full model output.

For a deeper look at Fable 5’s capabilities, pricing, and use cases, see the Claude Fable 5 guide.

The Mythos class explained

Anthropic’s Claude model lineup follows a four-tier hierarchy. The Mythos class, introduced with the Claude 5 generation, sits at the top:

Before the Claude 5 generation, Opus was the peak tier. The introduction of Mythos-class models restructured the hierarchy: Opus is now the premium-but-accessible tier, while Mythos represents the absolute capability ceiling.

The naming convention matters for understanding access. “Mythos 5” refers to the restricted model. “Fable 5” is the generally available Mythos-class model with added safeguards. Both are Mythos-class — the product name (Fable vs Mythos) signals the access level, not a difference in base intelligence.

How to get Mythos 5 access

Mythos 5 is not self-serve. You cannot select it in a model picker, call it through the public API, or access it through Claude.ai. There is no waitlist form on Anthropic’s website for general users.

Access is limited to:

1. Project Glasswing participants. Anthropic’s vetted program for organizations working on defensive cybersecurity, infrastructure protection, and approved biomedical research. Participation requires an existing relationship with Anthropic and approval through their trusted-access process.

2. Trusted-access partners. Organizations that Anthropic has directly onboarded for dual-use research. These are typically government-adjacent cybersecurity teams, critical infrastructure providers, and research institutions with appropriate oversight.

3. Cyber Verification Program (CVP) members. Members of the CVP get access to versions of Claude models with fewer cybersecurity restrictions. The CVP currently applies to Opus and Sonnet models; whether it extends to full Mythos 5 access or only to reduced-restriction Fable 5 variants is not publicly documented.

What if you do not qualify?

For the vast majority of users and organizations, the answer is straightforward: use Fable 5 or Opus 5. You get the same underlying model intelligence. The only difference is that certain high-risk requests will be blocked or rerouted — and for normal software engineering, analysis, writing, and business work, those classifiers rarely trigger.

If your work genuinely requires unrestricted cybersecurity or biology capability:

• Contact Anthropic’s enterprise or research partnerships team directly.
• Demonstrate a legitimate defensive or research use case.
• Expect a vetting process that evaluates your organization’s security posture, oversight mechanisms, and intended use.
• Do not expect rapid approval — this is intentionally a high-friction process.

Benchmark comparison: Mythos 5 vs Fable 5 vs Opus 5

Because Mythos 5 and Fable 5 share the same underlying model, their published benchmark scores are identical on non-safeguarded evaluations. The table below uses Anthropic’s published data from the June 9 launch and the July 24 Opus 5 announcement.

Where Mythos 5 diverges from Fable 5 and Opus 5

The benchmarks above do not capture the full Mythos 5 story because they avoid safeguarded domains. Based on Anthropic’s published statements:

• Cybersecurity — vulnerability discovery: On OSS-Fuzz, Opus 5 is close to Mythos 5 at finding vulnerabilities. Fable 5’s classifiers may reroute some of these requests to Opus 4.8, so the effective Fable 5 performance depends on whether the classifier triggers.

• Cybersecurity — exploit generation: Opus 5 is substantially behind Mythos 5 on exploiting discovered vulnerabilities. This is the clearest capability gap between the tiers. Fable 5’s classifiers block exploit generation entirely for general users.

• Long-running autonomous research: Mythos 5 is stronger than Opus 5 on extended autonomous research tasks. Opus 5 shows limitations on multi-hour research workflows where sustained, self-directed investigation is required.

• Biology and chemistry: Mythos 5 has the strongest raw capability in these domains. Fable 5’s biology safeguards are more restrictive, which is why Opus 5 (with fewer biology restrictions) is Anthropic’s most capable generally available model for scientific research.

Treat vendor-published benchmarks as directional. Validate on your own workload before committing budget.

Safety and safeguards: why Mythos 5 is restricted

The restriction on Mythos 5 is not arbitrary. Anthropic’s safety framework treats Mythos-class models as posing elevated risk in specific domains where advanced AI capability could enable real-world harm.

The dual-use problem

Mythos 5’s capabilities in cybersecurity, biology, and chemistry are strong enough to create a genuine dual-use dilemma:

• Cybersecurity: The same ability to find and exploit software vulnerabilities that helps cyberdefenders patch systems can help attackers compromise them.
• Biology: The same understanding of molecular biology and protein function that accelerates drug discovery could, in principle, inform the design of harmful biological agents.
• Chemistry: Similar dual-use concerns apply to advanced chemical reasoning.
• Model distillation: A sufficiently capable model could be used to train a smaller, less-aligned model that lacks safety classifiers entirely.

How Fable 5’s safeguards work

Fable 5 addresses this through classifiers that run alongside the model:

1. Detection: Classifiers analyze incoming requests for high-risk patterns in cyber, biology, chemistry, and distillation domains.
2. Intervention: Flagged requests are blocked, rerouted to Opus 4.8, or returned with summarized rather than full reasoning.
3. Conservative tuning: Anthropic openly acknowledges the classifiers are tuned conservatively, meaning some harmless requests trigger false positives.

In API contexts, a refused request returns HTTP 200 with stop_reason: "refusal" rather than a transport error. Production applications must handle this as normal control flow.

Opus 5’s lighter touch

Claude Opus 5 occupies a middle ground: its cyber classifiers intervene approximately 85% less often than Fable 5’s. It allows vulnerability discovery in source code but blocks binary-based vulnerability scanning, penetration testing, and exploit generation. Its biology safeguards are similar to Opus 4.8, making it the most capable generally available Claude model for scientific research.

Why not just make Mythos 5 safe and release it?

That is essentially what Fable 5 is. The question is whether the classifiers can be made precise enough to block misuse without degrading legitimate work. Anthropic’s current answer: not yet, for the highest-risk domains. The conservative tuning creates enough false positives in cyber and biology work that researchers and cyberdefenders need the unrestricted version — hence the restricted-access Mythos 5.

This is an active area of development. Anthropic’s AI alignment research aims to make future models safe enough for broader release without capability-restricting classifiers.

When Mythos 5 matters vs when Fable 5 or Opus 5 is enough

For most users reading this guide, the practical answer is: you do not need Mythos 5. Here is the decision framework.

Mythos 5 is relevant if:

• You are an approved cyberdefender who needs unrestricted vulnerability discovery and exploit generation
• You are an approved biology researcher working on tasks that Fable 5’s classifiers consistently block
• You are a critical infrastructure provider with a vetted defensive security program
• You need long-running autonomous research capability beyond what Opus 5 can sustain

Fable 5 is enough if:

• You need the highest generally available capability for autonomous coding, long-context analysis, or complex knowledge work
• Your work occasionally touches cyber-adjacent topics but does not require exploit generation
• You can accept 30-day data retention and occasional classifier false positives
• You want the strongest model on Claude.ai, Claude Code, or the public API

Opus 5 is enough if:

• You need near-Fable 5 performance at half the price ($5/$25 per MTok)
• You need zero data retention (Fable 5 and Mythos 5 both require 30-day retention)
• Your work is cybersecurity-adjacent and you want 85% fewer classifier interventions than Fable 5
• You need the freshest knowledge cutoff (May 2026 vs January 2026 for Fable 5)
• You are doing scientific research and want fewer biology restrictions than Fable 5

The full Claude model lineup in July 2026

For the detailed Opus 5 breakdown, see the Claude Opus 5 guide. For the Sonnet 5 breakdown, see the Claude Sonnet 5 guide. For the previous-generation comparison, see Fable 5 vs Opus 4.8. For the competitive landscape, see GPT-5.6 Sol vs Claude Fable 5.

FAQ

Bottom line

Claude Mythos 5 is the most capable model Anthropic has built — and the one you almost certainly cannot use. That is by design. The Mythos-class capability that makes it exceptional in cybersecurity, biology, and chemistry research is the same capability that makes unrestricted release irresponsible.

For everyone else, the good news is that you do not need it. Claude Fable 5 gives you the same underlying model intelligence through public channels. Claude Opus 5 gets you within half a percent of Fable 5 on coding benchmarks at half the price, with zero data retention and a fresher knowledge cutoff. The Claude model lineup in July 2026 is deep enough that the restricted tier is a footnote for most users — not a gap.

If you are a cyberdefender, infrastructure provider, or approved researcher who genuinely needs unrestricted Mythos-class capability, the path runs through Anthropic’s trusted-access programs. For everyone else: start with Opus 5, escalate to Fable 5 when the task demands it, and let Mythos 5 stay where it belongs — behind the glass.