Claude Fable 5 and Claude Mythos 5 Deep Dive — Anthropic's New "Mythos-Class" Top Tier Announced June 9, 2026, at $10/$50 per 1M Tokens, With Auto-Fallback to Opus 4.8 and Project Glasswing Gating

On June 9, 2026, Anthropic announced Claude Fable 5 and Claude Mythos 5. Both derive from the same base model but differ in safeguard configuration and access:

- Claude Fable 5 — public-facing, safeguards active, with auto-fallback to Opus 4.8 on high-risk queries - Claude Mythos 5 — same base, some safeguards loosened, restricted to Project Glasswing-vetted cyber defenders and biomedical researchers

The naming makes the intent explicit. Fable (Latin *fabula* — "what is told," i.e. publishable) and Mythos (Greek — "myth, archetype," the deeper, less constrained form). Both belong to a new "Mythos class" that sits above the existing Opus / Sonnet / Haiku tiers.

Source: Anthropic — Claude Fable 5 and Claude Mythos 5

This column is the follow-on to our Claude Code Agent View deep dive and Claude Code Automations / Routines column, focused on how Anthropic's tier structure has shifted and what that means for Japanese enterprises.

The new Mythos class sits on top of the existing tier ladder.

Fable = "the public storytelling face"; Mythos = "the underlying archetype, less constrained." The split maps directly to safeguard configuration.

Pro / Max / Team / Enterprise plans get Fable 5 bundled at no extra cost until June 22, 2026; from June 23 onward, usage-credit pricing applies.

Sources: Claude Fable, TechCrunch — Claude Fable 5 release

Fable 5's distinguishing operational feature: when high-risk queries are detected, the response is automatically served by Claude Opus 4.8 instead. Mythos 5 loosens these, but access itself is gated through Project Glasswing's review.

Three trigger categories:

1. Cybersecurity — attack-method generation or active exploitation aid 2. Bio-chemical — weapons of mass destruction, chemical weapons context 3. Distillation attempts — trying to extract training signal from Fable 5 itself

Operational notes:

- No Fable-tier billing when fallback fires; the request is served by Opus 4.8 - Fires on under 5% of sessions per Anthropic's disclosure - 1,000+ hours of external red-team and bug-bounty testing; no universal jailbreak found - 30-day data retention is specifically for safeguard monitoring

Sources: IT Pro — Mythos-class safeguards, CyberScoop — Mythos on a leash

Anthropic's framing: state-of-the-art on "nearly every AI capability benchmark tested," especially software engineering, knowledge work, vision, scientific research, and long-running tasks. Some benchmarks show Fable 5 outpacing Opus 4.8 by 10%+.

Third-party highlights:

- Hex core analysis benchmark: hit 90% for the first time ever - Genspark UI design / game code evaluation: best across all models - Cognition FrontierCode (our column): Claude Opus 4.8 currently leads at Diamond 13.4%; Fable 5 / Mythos 5 not yet on the public leaderboard

Sources: Inc., Yahoo Finance

Project Glasswing is the multi-stakeholder defensive-cybersecurity and biological-research initiative Anthropic launched in April 2026. Mythos Preview was supplied to ~50 partners and surfaced over 10,000 critical and high-severity vulnerabilities in critical-infrastructure software.

Mythos 5 is its successor with tight access conditions:

- Restricted to cyber defenders and biomedical researchers - 30-day data retention consent required - ID verification and use-case review by Anthropic - No direct contracting — Glasswing route only

Sources: Claude Mythos, Bleeping Computer — Mythos-class rollout

Source: Claude API Docs — Models overview

1. The tier hierarchy has changed — make sure proposals reflect it. Calling Opus "the top tier" is now out of date. Mythos class is the new flagship.

2. About 2× the price of Opus 4.8 — scope the use case. $10 / $50 per 1M tokens is roughly 3–4× Sonnet 4.6 ($3 / $15) and 10× Haiku 4.5 ($1 / $5). Restrict Fable 5 to highest-difficulty reasoning, long-horizon autonomy, code synthesis where top quality is required.

3. Understand the fallback behavior before integrating. Cyber / bio / distillation queries are silently routed to Opus 4.8. Differences in response style, performance, and context window will be observable. The 5% trigger rate may be much higher in pen-testing, drug-discovery research, or medical data analysis — possibly with false positives. Validate against your domain.

4. 30-day data retention deserves a governance review. Confidential information, PII, trade secrets — all are subject to the safety-monitoring retention window. Under Japan's amended Personal Information Protection Act, Economic Security Promotion Act, and sector-specific guidance (finance / healthcare / public sector), this should be reviewed with legal/compliance and locked into the DPA.

5. Mythos 5 is gated. Don't write "we use Mythos 5" if you don't. Only Fable 5 is directly contractable. Marketing materials and proposals claiming Mythos 5 usage without Project Glasswing membership are factually incorrect.

Anthropic's differentiation is concentrated in safeguard seriousness, Project Glasswing tie-in, and proven cyber-defense / biomedical impact (the 10,000+ critical vulnerabilities found in Preview).

As of June 10, 2026, the following remain officially undisclosed:

- Context window size (extension beyond Opus 4.8) - Fable 5 / Mythos 5 scores on [FrontierCode](../columns/cognition-frontiercode-benchmark-2026) — only Opus 4.8 is on the public leaderboard - Tokyo-region inference endpoint — US-only inference is documented, JP not yet - Computer Use (e.g. desktop OS control like Codex on Windows) availability under Fable 5 — not in the docs - Deprecation schedule for Opus 4.7 / 4.8 — only Opus 4.1's August 5, 2026 API retirement is on the books

Verify in the official model docs and your DPA before commitment.

Q1. Are Fable 5 and Mythos 5 literally the same model? A. Same base model. What differs is safeguard configuration and access. Fable 5 ships publicly with guardrails on; Mythos 5 has some guardrails loosened and is gated to Project Glasswing-vetted users. Q2. My workflow looks like it will trigger fallback constantly. What do I do? A. Pen-testing, drug discovery, clinical data — yes, false positives are plausible. Contact Anthropic support, describe the workflow, and ask about scoped exemptions. Or, intentionally use Opus 4.8 directly. Q3. How does a Japanese enterprise gain Mythos 5 access? A. Cyber defenders and biomedical researchers only, via Anthropic Japan or Project Glasswing. ID verification, use-case review, 30-day retention consent all required. Q4. Is $10 / $50 fair? A. It's under half of Mythos Preview and roughly 2× Opus 4.8. Justified for hardest reasoning / long-autonomy / SOTA code synthesis; for ordinary work, Sonnet 4.6 ($3 / $15) remains the cost-leader. Q5. The 30-day retention is a deal-breaker. A. Then stay on Sonnet 4.6 / Haiku 4.5, or open a Zero Data Retention (ZDR) conversation with Anthropic Japan under your Enterprise contract. Fable 5 / Mythos 5's safety-monitoring window may or may not be removable contractually — confirm before committing. Q6. AWS Bedrock / Google Vertex AI / Microsoft Foundry — all available? A. Yes, all three clouds. Model identifier `claude-fable-5`. Q7. Can I use Fable 5 from Claude Code ([Agent View](../columns/claude-code-agent-view-parallel-orchestration-2026))? A. The docs we've found are limited at this moment. Claude Code generally supports all Anthropic models, so presumably yes; check the latest plan limits and ACU equivalence in Anthropic's pricing pages. Q8. What changes June 22 / June 23? A. Bundled free in Pro / Max / Team / Enterprise plans through June 22; usage-credit pricing from June 23 onward. Get your finance team aligned before then.

Through our AI consulting practice, with FDE-style on-site enablement, we recommend the following pattern by default:

1. Haiku 4.5 — chat reply, summary, lightweight or real-time work 2. Sonnet 4.6 — business RAG, document generation, primary internal agent 3. Opus 4.8 — heavy reasoning, cross-repo analysis, long-text coherence checks 4. Fable 5 — flagship deliverables, benchmark-grade outputs, top-tier proposals, final-mile QA 5. Mythos 5 — cyber defense, biomedical research; direct application only

Claude Fable 5 / Mythos 5 marks a structural change in Anthropic's lineup: a new top tier above Opus, paired with safeguard-class differentiation between the public model and its less-restricted twin. The combination of safeguards, Project Glasswing gating, and 30-day retention is Anthropic embedding its safety stance directly into product tiers.

For Japanese enterprises, the five concrete things to lock in are: (1) reflect the new hierarchy in proposals, (2) ~2× pricing requires scoping, (3) understand the fallback behavior before integration, (4) governance review of 30-day retention, (5) Mythos 5 is invite-only — never claim usage you don't have. Internalizing these is the foundation for using Claude well from June 2026 onward.

Primary (Anthropic): - Claude Fable 5 and Claude Mythos 5 announcement - Claude Fable - Claude Mythos - Claude API Docs — Models overview - Claude Opus 4.7 announcement Third-party: - TechCrunch — Claude Fable 5 release - CNBC — Anthropic releases Mythos-class AI model - IT Pro — Fable 5 safeguards - CyberScoop — Mythos on a leash - Inc. — Fable 5 release - Yahoo Finance — Mythos-class Fable 5 - Bleeping Computer — Mythos-class rollout - Axios — Opus 4.8 release Related columns: - Claude Code Agent View deep dive - Claude Code Automations / Routines - Cognition FrontierCode benchmark - Windsurf × Devin integration - Google NotebookLM Better Research - Apple AFM Core Advanced - OpenAI Codex Computer Use on Windows - Forward Deployed Engineer (FDE) Note: Fable 5 and Mythos 5 share an identical base model and differ in safeguard configuration. As of June 10, 2026, context window size, a Tokyo region endpoint, Computer Use availability under Fable 5, and the deprecation schedule for Opus 4.7 / 4.8 are not officially documented. Mythos 5 is Project Glasswing invite-only — a general enterprise cannot contract for it directly.