Anthropic Unveils Claude Fable 5 and Mythos 5: Layered Release of Powerful, Risky AI

1. One underlying model, two names

Anthropic released Claude Fable 5 for all users and Claude Mythos 5 for trusted partners; both share the same base model but differ in safety guardrails. Fable 5 falls back to Opus 4.8 on high‑risk topics, while Mythos 5 removes some restrictions for vetted network‑defense and life‑science users.

2. Capability: Anthropic’s strongest publicly available model

Fable 5 is positioned as Anthropic’s most capable model for broad use, excelling in software engineering, knowledge work, vision, and scientific research, especially on longer and more complex tasks, making it a core for long‑task agents.

3. Programming: From code snippets to whole‑repo migrations

In a Stripe test, Fable 5 migrated a 50‑million‑line Ruby codebase in one day, a task that would take a team over two months manually. FrontierCode’s “medium effort” benchmark also gave Fable 5 the highest score among frontier models while using fewer tokens.

4. Vision, memory and knowledge work upgrades

Fable 5 achieves state‑of‑the‑art performance on complex visual tasks, extracting precise numbers from scientific charts and reconstructing web‑app source code from screenshots. In a “Pokémon FireRed” challenge it completed the game with minimal visual input. Long‑context memory improvements let it retain persistent files, tripling performance and final‑level reach compared with Opus 4.8 in a Slay the Spire test.

5. Mythos 5: Full capability for trusted scenarios

Mythos 5 is not a general stronger version but a version with some safety guardrails lifted for trusted partners. In Project Glasswing it provides enhanced network‑security abilities, and in life‑science workflows it accelerates protein‑design pipelines by roughly tenfold, handling binding‑site selection, tool execution, and failure recovery without human assistance.

6. Why the caution: Dual‑use capabilities are imminent

Fable 5’s safety classifiers route requests involving cybersecurity, biochemistry, or model distillation to Opus 4.8, with over 95 % of sessions not triggering a fallback. This reflects the growing dual‑use risk where the same capability can aid defenders or attackers.

7. Security: Harder to jailbreak, not just denied

Red‑team testing (over 1,000 hours) found no generic jailbreaks, and external partners reported stronger protection against harmful cyber queries than older models. The approach prefers model fallback rather than outright refusal to preserve user experience while narrowing false‑positive risk.

8. Biological capability: Benefits and risks

For most biochemistry queries Fable 5 falls back to Opus 4.8, but Mythos‑class models can predict AAV capsid properties better than specialized protein‑language models despite no explicit training, illustrating the double‑edged nature of advanced AI in life‑science.

9. Data retention policy

All traffic to Fable 5, Mythos 5, and future Mythos‑class models is retained for 30 days, not used for training new Claude models, and is deleted after the retention period, posing compliance considerations for enterprise users.

Fable 5 and Mythos 5 represent more than a performance jump; they mark Anthropic’s shift to a layered capability release where ordinary users receive a guarded strong model, trusted professionals obtain fuller power, and the platform manages risk through classifiers, fallbacks, trusted access, and data audit.