Claude Fable 5 and Mythos 5: Capabilities

Abstraction: Fable 5 capability review, benchmarks, classifiers, reception

Key points:

• Fable 5 is a Mythos-class Anthropic model made "safe for general use," SOTA on nearly all benchmarks, priced $10/$50 per M tokens (2x Opus), 30-day data retention required; taken offline by the US government 3 days after release.
• Benchmarks: Cursor Bench 72.9% (vs GPT-5.5 64.3%), GPQA Diamond 94% (saturated), USAMO 2026 99.8%, RiemannBench 55%, FrontierMath Tiers 1-3 at 87%; #1 on Agent Arena by widest margin ever, but weaker steerability.
• Safety classifiers auto-route cybersecurity/biology-chemistry/distillation requests to Opus 4.8; tuned to avoid false negatives at cost of absurd false positives (triggered by "cancer," "hi," panspermia papers); any usable LLM can be jailbroken.
• Reception overwhelmingly impressed: Karpathy calls it a major step-change; Boris Cherny cites "big model smell"; Taelin reports Fable achieving a 1770% speedup and finding a subtle bug he didn't ask about; Simon Willison notes it hacks together unrequested browser-screenshot tooling.
• Strong on coding, fiction/prose (Yudkowsky, Kendric Tonn), math (economists Josh Gans, Vincent Grégoire found closed-form proofs); classifier friction pushed some power users (SemiAnalysis) toward OpenAI Codex.

Connections: Zvi Mowshowitz · Anthropic · Claude · Andrej Karpathy · Frontier AI · AI Safety Classifiers · Coding Assistants

Source: https://thezvi.substack.com/p/claude-fable-5-and-mythos-5-capabilities