The weird situation with Fable

TL;DR

• 1

  Fable 5 is basically Mythos 5 with guards at the door: Theo stresses that Fable and Mythos 5 are the same underlying model, but Fable adds aggressive classifiers that reroute some requests to Opus 4.8 or block them entirely.

• 2

  Anthropic initially hid AI-research restrictions from users: For frontier LLM development prompts, Anthropic said it could use prompt modification, steering vectors, or parameter-efficient fine-tuning to limit effectiveness without telling the user, affecting about 0.03% of traffic and fewer than 0.1% of orgs.

• 3

  The enterprise data policy is a huge blocker: Anthropic now requires 30-day retention for Mythos-class traffic, and flagged policy-violation chats can keep inputs and outputs for up to two years and trust-and-safety scores for up to seven years, which Theo says rules out many Fortune 500 use cases.

• 4

  False positives make the user experience messy and expensive: Theo shows harmless requests, like help with a Defcon Gold Bug cryptography puzzle, getting rerouted to Opus and then failing anyway, while users may be billed at Opus rates.

• 5

  Anthropic changed the system card after release: Mid-recording, Theo notices the public system card had been swapped with a revised version, which he frames as Anthropic trying to rewrite history around the most controversial safeguard language.

• 6

  The deeper fear is trust erosion: Theo argues that once a model can secretly degrade outputs on sensitive competitive tasks, developers can no longer tell whether a bad answer came from model limits, bad context, or an invisible policy intervention.