MYTHOS is LIVE!!!!

TL;DR

• 1

  Fable 5 is the public, guarded version of Mythos: Berman frames Mythos as the security-community model and Fable as the same Mythos-class system with added guardrails for general users.

• 2

  The benchmark jump is big, especially for coding: Anthropic claims SWEBench Pro at 80% for Mythos/Fable versus 69% for Claude Opus 4.8 and 58% for GPT-5.5, plus 88% on Terminal Bench versus 83.4% for Opus 4.8.

• 3

  Real-world behavior felt more important than the charts: Berman says Fable "feels bigger" than other models, eagerly exploring entire codebases and handling long-horizon tasks without stumbling, even when he tried to find something too complex for it.

• 4

  The biggest annoyances were speed, verbosity, and constant clarification loops: In his testing, Fable would ask 3 to 5 follow-up questions, summarize his answers, ask to write a spec, ask to confirm the spec, then ask whether to use parallel agents before finally starting work.

• 5

  Pricing makes model routing mandatory: At $10 per million input tokens and $50 per million output tokens, Berman argues Fable should be saved for the hardest tasks while Sonnet or Haiku handle everything else.

• 6

  Live demos showed both the ceiling and the cost: On stream, Fable built a realistic 3D Rubik’s Cube that scrambled and solved correctly, plus a striking browser-based fluid simulation, while burning through 51% of his current Claude Max 20x session quota.