Claude Fable 5

Which lens this is. The model-as-substrate treatment: what Fable 5 costs and constrains for the LLM-maintained-wiki thesis, plus the guardrail and export-suspension arc. Its market standing — list price, positioning against Opus 5 / Sonnet 5, third-party benchmarks — is a separate page, llm-providers-wiki/claude-fable-5. Same slug in two spokes is deliberate (HUB.md → dual-lens); link the lens you mean explicitly.

anthropic‘s frontier LLM released 9 June 2026, introduced by claude-fable-5-review (Simon Willison) and confirmed by the primary Anthropic announcement. The newest point in the model-substrate thread, successor in role to claude-opus-4-8.

What we know

• A guardrailed sibling — now confirmed, with the mechanism. Fable 5 and claude-mythos-5 are the identical underlying model; the safeguards are the only difference (Anthropic: “Fable… ‘that which is told,’ akin to Greek mythos. The safeguards are what distinguish the two models”). Fable 5 is “a Mythos-class model made safe for general use” via three safety classifiers — cybersecurity (block offensive-cyber/exploit-dev), biology/chemistry (route most biomedical queries to claude-opus-4-8 for dual-use risk), distillation (block capability extraction). Crucially it’s not a weaker model — it’s Mythos 5 + a fall-back layer, and >95% of sessions involve no fallback at all. So the “safety-tier split” is capability parity, gated by deployment policy, not a capability cut.
• Specs (from the review; not restated in the announcement): 1M-token context, 128K max output, Jan 2026 knowledge cutoff.
• Pricing: $10 / $50 per Mtok — 2× claude-opus-4-8 ($5/$25), but Anthropic frames it as “less than half the price of Claude Mythos Preview” (both reference points hold). Free on subscriptions 9–22 June 2026, then usage credits; API + consumption Enterprise; staged rollout. Independently corroborated (noted 2026-08-03): llm-providers-wiki/claude-fable-5 reached the same $10/$50 from three unrelated third-party reads plus arithmetic on Anthropic’s own Opus 5 post, without ever seeing this announcement. Four independent routes to the same number is unusually strong for a pricing claim in this corpus — and worth noting because that spoke recorded the figure as unconfirmed the whole time the primary source sat here.
• Capability signals: Willison — “substantially larger,” deeper knowledge (lists his projects with accurate dates), MicroPython→CPython-WASM port. Anthropic narrative claims — Stripe’s 50M-line Ruby migration in a day (vs two months), top score on Hebbia’s finance benchmark, vision SOTA (finished Pokémon FireRed by vision alone). Benchmarks are qualitative in the announcement (no numeric SWE-bench/GPQA given).
• New data policy: 30-day retention on all Mythos-class traffic (safety-only, access logged).

Temporary suspension — June 2026 (claude-fable-5-infoq)

A US government export directive pulled Fable 5 offline within three days of launch. Amazon’s security team identified a jailbreak vulnerability and escalated to the White House; david-sacks (White House AI adviser) said publicly the suspension was expected to be temporary, pending Anthropic remediating the issue.

The 30-day retention policy also created cross-vendor friction: Microsoft’s standard enterprise contracts require zero retention, which conflicts with the requirement both Fable 5 and claude-mythos-5 carry.

The government directive is a governance action — cross-spoke context for ai-governance-wiki.

Resolved — redeployed 1 July 2026 (redeploying-fable-5, primary). Anthropic’s own post gives the precise mechanism and the end of the arc: the 12 Jun US export controls hit both Fable 5 and claude-mythos-5, and Anthropic suspended both because real-time nationality verification of foreign nationals was impossible (not a partial restriction). Controls were lifted by 30 Jun; Fable 5 redeployed 1 Jul (Claude Platform / Claude.ai / Claude Code / Cowork; up to 50% of weekly limits through 7 Jul, then credits; cloud to follow). Two substantive nuances: (1) the Amazon-found jailbreak let Fable 5 identify/exploit software vulns, but Anthropic’s testing found “many less capable models—including claude-opus-4-8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities,” so the exposed capability was not unique to Fable 5; (2) the cyber classifier was hardened — it now blocks the reported behavior in >99% of cases with a deliberate “safety margin,” still falling back to claude-opus-4-8. Anthropic also launched a multi-vendor jailbreak-severity framework (with Amazon/ Microsoft/Google) + a HackerOne cyber-jailbreak program — governance material for ai-governance-wiki.

Role in this wiki

Same substrate role as claude-opus-4-8: the model class beneath the agent tooling (agentic-tooling-wiki, cross-wiki) and the Claude-driven knowledge bases here (gbrain, llm-wiki-agent). Two threads it touches:

• The “and cost” half. Fable 5 doubles the per-token price of Opus 4.8 — the substrate is getting more capable and more expensive at the frontier, which raises the cost pressure on the “near-free maintenance” bet the llm-wiki/gbrain thesis leans on (cf. anthropic-valuation‘s unproven-economics flag). The cheaper-substrate trend (Opus 4.8 had unchanged pricing vs 4.7) does not continue here.
• Honesty/safety. Where claude-opus-4-8‘s headline was honesty, Fable 5’s is guardrails (vs the looser Mythos 5) — the safety story moving from truthfulness to restriction. Worth watching whether the heavier guardrails reproduce the “neurotic over-refusal” reception claude-opus-4-8-zvi flagged for 4.8.

For cross-market standing this model would also belong on llm-benchmarks / llm-api-pricing (llm-providers-wiki, cross-wiki) — the runner-up spoke for its source.

Related

claude-fable-5-review · claude-fable-5-infoq · redeploying-fable-5 · claude-opus-4-8 · anthropic · gbrain · llm-wiki-agent · anthropic-valuation · david-sacks

Sourced from a hands-on review (claude-fable-5-review) and now the primary Anthropic announcement (claude-fable-5-mythos-5-announcement). Post-dates the wiki’s knowledge cutoff; pricing/availability are dated snapshots; announcement benchmarks are qualitative/vendor-framed. Maker: anthropic.