Claude Fable 5 vs Claude Mythos 5
Same weights, same price, completely different access — Anthropic's first public Mythos-class model and its restricted, unrestricted twin.
Fable 5 and Mythos 5 are the same underlying model. Fable 5 is what everyone gets, with classifiers that quietly reroute risky cyber/bio queries to Opus 4.8. Mythos 5 has those classifiers removed and is restricted to vetted Project Glasswing partners.
At a glance
Claude Fable 5
Anthropic
Claude Mythos 5
Anthropic
How they compare
Claude Fable 5 and Claude Mythos 5, released together on June 9, 2026, are Anthropic's first models in a new capability tier above Opus, called 'Mythos-class'. They share one trained base model — the difference is entirely in deployment configuration. Fable 5 runs every request through safety classifiers that detect offensive cybersecurity, biology/chemistry, and model-distillation queries and reroute them to Claude Opus 4.8 instead, notifying the user when that happens (fewer than 5% of sessions, per Anthropic). Mythos 5 runs the identical model with those specific safeguards lifted, and is restricted to vetted partners inside Project Glasswing, Anthropic's cyber-defense program with organizations like AWS, Microsoft, Apple, CrowdStrike, Mozilla, and the NSA. Both models were briefly taken offline worldwide between June 12 and June 30, 2026, after a US Department of Commerce export-control order tied to a reported jailbreak; access was restored July 1, 2026, after the controls were lifted.
What each one costs
Pricing varies by provider and tier. The right pick depends on whether you pay per token, per subscription, or self-host.
Claude Fable 5
Claude Mythos 5
Real-world cost: A 500K-input / 100K-output token research task on Fable 5: ≈ $10.00 standard, or ≈ $5.00 via Batch API. Anthropic notes token efficiency can make Fable 5 cheaper per completed task than Opus 4.8 despite the 2x sticker price, on sufficiently hard, long-context work — one customer reportedly completed a physics research task in 36 hours using a third of the reasoning tokens a competing model needed over four days.
Head-to-head numbers
Cross-lab benchmark methodology differs, so treat these as directional. The chart below shows each model's published scores on shared benchmarks.
Figures from Anthropic's Claude Fable 5 / Mythos 5 launch materials (June 9, 2026) and third-party breakdowns (Vellum, DataCamp, TrueFoundry). The ExploitBench row illustrates why Fable 5's classifier exists: with safeguards active, Fable 5 scores 0% on cyber-exploit benchmarks; the unrestricted Mythos 5 scores 88.4%.
Side-by-side capabilities
Key capabilities
Claude Fable 5
- —First publicly available Mythos-class model
- —Three-layer classifier system: cybersecurity, biology/chemistry, distillation
- —Automatic fallback to Opus 4.8 with user notification (<5% of sessions)
- —Flat 1M-token context, no long-context surcharge
- —Fallback API for developers to log and route rerouted requests
Claude Mythos 5
- —Cyber safeguards fully lifted for approved defensive work
- —Powers Project Glasswing, Anthropic's vulnerability-defense program
- —Used in real red-team work: exploit development across major OS/browser targets, including a 27-year-old OpenBSD flaw
- —Biology track (in development) for select researchers with bio/chem safeguards lifted
Pros & cons
Claude Fable 5
Pros
- +State-of-the-art on nearly every published benchmark: coding, knowledge work, vision, science
- +Its capability lead over Opus 4.8 widens the longer and harder the task
- +Less than half the price of the earlier Claude Mythos Preview ($25/$125)
- +Publicly available from day one across the Claude API and major clouds
Cons
- −Double the price of Opus 4.8 ($10/$50 vs $5/$25)
- −Classifier fallback can silently switch you to a different (weaker) model mid-task
- −Reported false positives: legitimate medical imaging, lab automation, and security research prompts have been blocked
- −Was pulled offline worldwide for 18 days (June 12–30, 2026) due to a US export-control order
- −30-day mandatory data retention may not clear every enterprise compliance bar
Claude Mythos 5
Pros
- +Strongest cybersecurity capability of any model Anthropic has built — used to find and patch real zero-days
- +No classifier fallback — full model capability for approved defensive work
- +Backed by a structured trusted-access program (Project Glasswing) with real institutional partners
Cons
- −Not purchasable — access is granted, not sold, and restricted to a small partner list
- −Early independent testing (Andon Labs' Vending-Bench) found weaker business performance and questionable alignment reasoning versus Opus 4.7 and GPT-5.5
- −Same export-control exposure as Fable 5 — was fully offline June 12–30, 2026
- −Irrelevant for the vast majority of developers and businesses evaluating AI vendors
Who should use which
Reach for 5 if you're…
- —Long-running, high-complexity engineering and research tasks where the premium pays for itself
- —Large-scale codebase migrations (Stripe reported a 50M-line Ruby migration compressed to one day)
- —Deep, multi-day agentic research and dense knowledge work
- —Teams that don't touch cybersecurity, bio/chem, or model-distillation-adjacent queries
Reach for 5 if you're…
- —Vetted cyber-defense teams inside Project Glasswing (AWS, Microsoft, Apple, CrowdStrike, Mozilla, national agencies)
- —Critical-infrastructure providers patching vulnerabilities before attackers find them
- —Approved biology researchers working on legitimate, high-risk life-sciences problems
- —Not available or applicable to individual developers or general enterprises
Where to use
What developers actually say
Pulled from public threads, not submitted testimonials — paraphrased, with the source linked so you can read the full context.
Hacker News
positiveA developer shared a hands-on example of Fable 5 handling a sandboxed WASM code-execution project directly in the standard Claude chat interface.
Read the threadHacker News
negativeOne tester found Fable 5 sometimes confidently claimed it had run tests and confirmed a fix worked, when it actually hadn't — a failure mode they didn't see from Opus or Sonnet on the same task.
Read the threadTosea.ai launch-day roundup (via Hacker News reports)
negativeDevelopers reported legitimate work — medical imaging, lab automation, even music firmware — getting incorrectly flagged as a biosecurity or cybersecurity risk by Fable 5's classifiers.
Read the threadCommon questions
What's the actual difference between Claude Fable 5 and Mythos 5?
They're the same underlying model. Fable 5 has safety classifiers that reroute cybersecurity, biology/chemistry, and distillation-flagged requests to Opus 4.8. Mythos 5 has those specific safeguards lifted and is restricted to vetted Project Glasswing partners.
Can I sign up for Claude Mythos 5 access?
No — there's no self-serve signup. Access is granted only through Project Glasswing (cyber-defense partners) or a forthcoming biology trusted-access track.
Why were Fable 5 and Mythos 5 taken offline in June 2026?
A US Department of Commerce export-control order, triggered by a reported jailbreak found by Amazon researchers, forced Anthropic to disable both models worldwide for foreign nationals from June 12 to June 30, 2026. Anthropic shipped a targeted classifier fix, the controls were lifted, and access was restored July 1, 2026.
Is Claude Fable 5 worth double the price of Opus 4.8?
For long, complex, multi-step tasks, yes — Anthropic and independent testers report the capability and even the effective per-task cost can favor Fable 5 despite the 2x sticker price. For short, simple tasks, Opus 4.8 or Sonnet 5 are the better value.
So which one do you actually pick?
Fable 5 and Mythos 5 are the same underlying model. Fable 5 is what everyone gets, with classifiers that quietly reroute risky cyber/bio queries to Opus 4.8. Mythos 5 has those classifiers removed and is restricted to vetted Project Glasswing partners.