Restored for approved US organizations — July 1, 2026
Mythos 5 access has partially resumed
The June 12, 2026 export-control directive took Mythos 5 offline along with Fable 5. On June 26 the US government approved restoring Mythos 5 for a set of trusted US organizations, and access resumed July 1. Anthropic is working to expand access to broader domestic and international Project Glasswing partners — there is still no public route.
The full suspension & restoration story →Restored for approved US organizations — July 1, 2026
Claude Mythos 5: the restricted model behind Fable5
Claude Mythos 5 is Anthropic's most capable model — the same underlying model as Claude Fable5, with the cybersecurity and biology safeguards lifted. It is not publicly available: access is limited to vetted cyberdefenders and critical-infrastructure providers through Project Glasswing. Here is everything verifiable about it — and what you can use instead.
Claude Mythos 5 vs Fable 5
One model, two releases. The differences are entirely about safeguards, access, and terms — not capability.
| Dimension | Claude Mythos 5 | Claude Fable5 |
|---|---|---|
| What it is | The unrestricted Mythos-class model | The same model with safety classifiers added |
| Safeguards | Lifted — full capability in cybersecurity & biology domains | Classifiers reroute flagged requests to Opus 4.8 (fewer than 5% of conversations) |
| Access | Project Glasswing — vetted organizations only | Generally available, no waitlist |
| Channels | Direct Anthropic engagement | Claude apps, API, Bedrock, Google Cloud, Foundry, Copilot, OpenRouter |
| Pricing | Not public — negotiated per engagement | $10 / $50 per million tokens |
| Data retention | 30-day safety retention | 30-day safety retention; ZDR agreements do not apply |
| Published benchmarks | The starred rows — cyber & bio evals like ExploitBench | Everything else in the June 9 table |
Sourced from Anthropic's June 9, 2026 announcement and the shared system card. We update this table when the policy changes.
Reading the Mythos 5 benchmark rows honestly
Anthropic's launch table mixes the two models — the headline scores everyone quotes deserve a footnote.
01
The starred rows in Anthropic's benchmark table — cybersecurity and biology evaluations like ExploitBench — are Mythos 5 scores, not Fable5. They're real, but you almost certainly can't reproduce them.
02
Fable5's own results in those flagged domains land near Opus 4.8 levels — because when a safeguard fires, Opus 4.8 is literally the model answering.
03
Coding, knowledge work, and computer-use scores are genuine Fable5 results — for those, what's published is what you get.
Rule of thumb: if a domain is restricted enough to trigger Fable5's safeguards, the published top score belongs to Mythos 5 — and the score you'll actually experience belongs to Opus 4.8.
See the full Fable5 benchmark analysisWho gets access — and how
Anthropic kept Mythos 5 deliberately scarce. The published route is narrow and organizational — there is no individual access.
Project Glasswing
Anthropic's vetted-access program for Mythos-class capability. It covers cyberdefense teams and critical-infrastructure providers that pass organizational vetting.
Who qualifies
Organizations doing defensive security or infrastructure-hardening work — not individuals, not general enterprises, and not researchers outside the covered domains.
How to apply
Through direct engagement with Anthropic. There is no public application form, waitlist, or self-serve signup — if your organization qualifies, your Anthropic contact is the route.
No beta, no free tier
Despite what the 'mythos 5 beta' and 'mythos 5 free' searches suggest, neither exists. There is no public beta and no free access of any kind.
For everyone else, Claude Fable5 is the Mythos-class model you can actually use — the same underlying model, safeguarded, and generally available today.
How to access Fable5 (API, apps, clouds)Confirmed vs rumor
Mythos 5 generates more Reddit speculation than verifiable fact. Here is the current split — updated as claims get confirmed or debunked.
Confirmed (primary sources)
- Mythos 5 is the same underlying model as Fable5 with safeguards lifted, announced June 9, 2026 alongside Fable5.
- Access runs through Project Glasswing and is limited to vetted cyberdefenders and critical-infrastructure providers.
- The 30-day data retention policy covers Mythos 5 the same as Fable5, and the shared system card documents both variants.
Rumor / unverified
- Lists of specific companies with Mythos access circulate on Reddit — Anthropic has not published any customer list.
- Claimed 'Mythos 5 leaks' (weights or outputs) have surfaced in community threads; none have been substantiated.
- A 'Mythos 5.0 beta' opening to the public — no such program exists or has been announced.
Mythos 5 FAQ
The questions people actually search — answered from primary sources, with rumors labeled as such.
Build on Claude Fable5 with your eyes open
Know the Claude Fable5 cost math, the safeguard behavior, and the right tier before you commit. Create a free account to get notified when specs, prices, or access channels change.