Claude Fable 5 and Claude Mythos 5 (Anthropic announcement)

The primary Anthropic source the Willison review flagged as missing — the official 9 June 2026 launch of two Mythos-class frontier models, Anthropic’s most capable to date. It confirms the review’s “guardrailed sibling” framing and reveals the mechanism behind it.

The key revelation: same model, safety is the only difference

claude-fable-5 and claude-mythos-5 are the identical underlying model; the safeguards are the sole distinction (the name nods to this — fabula “that which is told” ≈ Greek mythos). Fable 5 is “a Mythos-class model made safe for general use”; Mythos 5 is the same model “with the safeguards lifted in some areas” for authorized users.

Fable 5’s safety layer = three classifiers that fall back to claude-opus-4-8:

1. Cybersecurity — blocks offensive-cyber / exploit-dev tasks.
2. Biology/Chemistry — routes most biomedical queries to Opus 4.8 (dual-use risk).
3. Distillation — blocks capability-extraction attempts. >95% of Fable sessions involve no fallback — so for ordinary use Fable ≈ Mythos.

Other primary facts

• Pricing (both): $10 / $50 per Mtok — Anthropic frames this as “less than half the price of Claude Mythos Preview” (note: it is 2× claude-opus-4-8; both reference points hold).
• Benchmark claims are narrative, not numeric: Stripe — “months of engineering into days,” a 50M-line Ruby migration in one day vs two months; highest score on Hebbia’s Finance Benchmark; vision SOTA, completed Pokémon FireRed “with vision alone”; Mythos 5 matched/beat skilled humans on protein design.
• Availability: Fable 5 everywhere today (free on subscriptions 9–22 June, then credits; API + consumption Enterprise), staged/“conservative” rollout. Mythos 5 restricted — Project Glasswing (US-government), cybersecurity partners now, a biology trusted-access program “in coming weeks.”
• New data policy: 30-day retention on all Mythos-class traffic, safety-only, access logged, deleted after 30 days in almost all cases.
• Alignment: Mythos 5 misaligned behavior “low, similar to Opus 4.8” (Fable presumed equal).

Not disclosed here: context window, max output, knowledge cutoff, numeric benchmarks — for those the wiki still relies on claude-fable-5-review (1M ctx / 128K out / Jan-2026 cutoff). Vendor announcement: positioning/benchmark framing are Anthropic’s own.

Related

claude-fable-5 · claude-mythos-5 · claude-fable-5-review · claude-opus-4-8 · anthropic