Fable is Mythos, and it is really good.

Claude's new Fable 5 model (publicly available version of Mythos 5) is reviewed after extensive hands-on testing involving over $2,000 in inference spend. The model is praised as the best coding model ever released, capable of modernizing large legacy codebases, generating complex games with multiplayer support, and writing higher-quality code with better 'taste.' Key concerns include aggressive safety restrictions that silently degrade model performance (prompt modification, steering vectors), mandatory 30-day data retention for all Fable 5 traffic, extremely high token costs ($100 in 8 minutes on usage-based billing), and session limits. Benchmark results show strong performance on Frontier Codebench and vision tasks, but significant score drops due to refusals on certain question types. The post encourages developers to push the model harder and rethink how they approach AI-assisted development.