Fable is Mythos, and it is really good.

TL;DR

• 1

  Fable 5 feels like Mythos 5 with a muzzle on: Theo says the public model is clearly powerful but loses benchmark points and real-world utility because safety layers trigger refusals, fallback routing, and silent performance limits on some topics.

• 2

  The coding jump is big enough that Theo calls it the best model yet: He compares GPT-5.5 to a rebuilt GPT-5.4, while Mythos feels like "more Opus turned up to 12 or 13," with better code quality, stronger UI work, and more initiative on vague tasks.

• 3

  The economics are wild, both good and bad: Fable costs $10 per million input tokens and $50 per million output tokens, spent $100 in about 8 minutes on usage-based billing, yet still often beats Opus on total cost because it uses fewer tokens and gets more done per run.

• 4

  Benchmarks are strong, but Theo trusts some more than others: He highlights an 80% on SWEBench Pro and a 30% on Frontier Code's hardest tier, but calls SWEBench Pro weak and says Frontier Code has suspicious variance, while DeepSWE lines up better with his hands-on experience and shows Fable near GPT-5.5.

• 5

  Theo's team stress-tested the model with real software, not toy demos: In one day they made a terminal-based 2.5D adventure in Rust, a full multiplayer 3D racing game with spectator mode, a Minecraft clone, and a Rust port of T3 Chat-style tooling.

• 6

  The tradeoffs are serious for companies: Fable requires 30-day retention for all traffic, cannot use Anthropic's no-retention setup, and may silently get weaker on frontier model development tasks via prompt modification, steering vectors, or parameter-efficient fine-tuning.