# Fable 5 / Mythos 5: One Model, Two Governance Surfaces—Then Export Controls Tested the Template

Times of AI Desk · 2026-06-09 · Models

[https://timesof.ai/2026/06/anthropic-claude-fable-5-mythos-5-launch](https://timesof.ai/2026/06/anthropic-claude-fable-5-mythos-5-launch)

> Anthropic split Mythos-class into Fable 5 (general, classifiers, Opus 4.8 fallback <5% sessions) and Mythos 5 (Glasswing trusted). AA Intelligence Index #1 at ~64.9; $10/$50. Dual-use release template that a June 12 Commerce freeze would immediately stress-test.

Fable/Mythos is Anthropic’s answer to releasing near-frontier dual-use capability without a full open dump: one model, two governance surfaces. It operationalizes the Glasswing thesis—general users get extreme capability under classifiers; defenders get full cyber horsepower under trust boundaries. The mid-June export-control interruption that followed only underscores how tightly these launches now couple product, safety, and geopolitics.

Anthropic split its strongest Mythos-class system into two products: **Claude Fable 5** for general release with new classifiers and fallbacks, and **Claude Mythos 5** with cyber (and later biology) safeguards lifted for vetted partners. Anthropic positions Fable 5 as exceeding any prior generally available Claude on software engineering, knowledge work, vision, long-horizon autonomy, and science—while acknowledging dual-use cyber and bio risks that forced Project Glasswing–style restricted preview until safeguards matured. Pricing: **$10 input / $50 output per million tokens**.

## Two products, one underlying model

**Claude Fable 5 (general)**
- Mythos-class model “made safe for general use.”
- State-of-the-art claims across tested capability areas; larger leads on long, complex tasks.
- **Safety classifiers** route sensitive topics to **Claude Opus 4.8** instead of Fable: cybersecurity, biology/chemistry, and distillation-related patterns.
- Average **<5% of sessions** trigger fallback; >95% run fully on Fable in early data.
- Conservative tuning means some benign requests may false-positive; Anthropic aims to narrow classifiers over time.
- **30-day retention** required for business traffic on Mythos-class models (safety uses only; not for training).

**Claude Mythos 5 (restricted)**
- Same base model with safeguards lifted in designated areas.
- Initial path via **Project Glasswing** for cyberdefenders/infrastructure providers (upgrade from Mythos Preview).
- Strongest claimed cyber capabilities of any model; biology trusted-access program planned with bio safeguards removed but cyber still on.
- Expanded trusted-access program intended later with U.S. government consultation.

## Capabilities and customer signals

Anthropic and early testers report large gains on long-horizon coding (e.g., Stripe multi-month migration compressed to a day in a 50M-line Ruby codebase), finance/document reasoning (Hebbia, IMC), vision (Pokémon FireRed with vision-only harness), memory-augmented gameplay, and scientific workflows (protein design acceleration, novel genomics ML). Alignment assessment: misaligned behavior rates similar to Opus 4.8. Full evaluation suite in the Fable/Mythos system card.

## Claims vs checks

| Source | Snapshot | Standing |
|--------|----------|----------|
| **Artificial Analysis Intelligence Index** | AA / LinkedIn AA posts **Jun 9, 2026** | **Fable 5 ≈ 64.9** — **#1 overall**, ~**5 points** clear of the best non-Anthropic model (GPT‑5.5 band); Anthropic models take **top 2** slots. Leads **5 of 10** underlying AA benches in launch writeups |
| **GDPval-AA** | AA Fable article **Jun 9** | **#1** agentic knowledge-work Elo (**~1932**); Anthropic holds **3 of top 4** on that board |
| **Arena Text Overall** | Mid-July boards (post-restore) | **claude-fable-5 #1** at **~1507 Elo** (Jul 16 Arena snapshot)—the durable preference leader after the Jun 12–30 export freeze and Jul 1 restore |
| **Arena Frontend Code** | Jul 16 board | Fable **#2** at **~1631** behind **Kimi K3 #1 / 1679**—still #1 closed model on that coding surface until K3 unblinded |

Vendor capability claims and the AA/Arena #1 stack are the independent half of the Fable story: Mythos-class general access that **actually took the public intelligence and chat preference crowns**.

## Pricing and availability

- **$10 / $50** per million input/output tokens for both Fable 5 and Mythos 5 (under half Mythos Preview pricing).
- API / consumption Enterprise: Fable fully available at launch (`claude-fable-5`).
- Subscriptions (Pro, Max, Team, seat Enterprise): included at no extra cost **through June 22**, then usage credits unless capacity allows extending; goal to restore standard plan inclusion when capacity allows.

*(Note: Access was later suspended June 12 under a U.S. export-control directive and subsequently redeployed—outside this launch-day story.)*

## Limits

- Classifier fallback <5% of sessions is early Anthropic data—false positives on benign cyber/bio queries remain.
- Customer anecdotes (Stripe, Hebbia, etc.) are lab-cited; not independent task audits.
- Arena #1 preference leadership in this table is mid-July post-restore—launch day itself was interrupted three days later.

## Sources

- Anthropic: [“Claude Fable 5 and Claude Mythos 5”](https://anthropic.com/news/claude-fable-5-mythos-5) (anthropic.com/news/claude-fable-5-mythos-5, June 9, 2026). Primary announcement with pricing, safeguards, availability, and capability claims.
- Claude Fable 5 / Mythos 5 system card (linked from announcement).
- Artificial Analysis: “Claude Fable 5: the first public Mythos-class model” and Intelligence Index #1 launch notes (June 9, 2026).
- Arena.ai Text Overall leaderboard snapshots (mid-July 2026) for post-restore preference leadership.
- CNBC and contemporaneous coverage of the dual-product Mythos-class launch (June 9, 2026).