What it feels like to work with Mythos
| Source: One Useful Thing (Ethan Mollick)
Tags: Claude Fable 5, Anthropic, Mythos, AI capabilities, benchmark, Ethan Mollick
Wharton professor Ethan Mollick's early-access test of Claude Fable 5 finds it outperforms every prior model by a wide margin — running autonomously for up to 12 hours on complex tasks — and describes the experience as 'delightful and unnerving.'
Details
Ethan Mollick, a Wharton professor and one of the most cited AI researchers in applied settings, was given early access to Claude Fable 5 before its public launch. His assessment is unambiguous: Fable consistently and substantially outperformed all other public models he tested it on, across academic writing, creative tasks, and engineering challenges. Notably, the model would autonomously execute multi-page specifications for up to 12 hours without further guidance. As a concrete demonstration, Mollick had Fable build several playable browser games from brief natural language prompts, with all visual assets generated mathematically rather than from external image files. His most technically revealing test involved isochrone maps — geographic visualizations of travel time that no prior model had handled well — which Fable executed correctly from scratch. He also describes the model producing a sophisticated academic social science paper and a 10-page epic poem with every word starting with 's' from single prompts. The piece warns that Fable's cybersecurity guardrails are strict enough to prevent testing in that domain. Mollick frames the overall experience as signaling a qualitative shift in human-AI collaboration rather than just another incremental improvement.