What it feels like to work with Mythos
Points and comments are a snapshot, not live.
Claude 5 Fable outperforms prior models across tasks but shifts user role from doing work to commissioning results.
Ethan Mollick tested Fable, Anthropic's latest model, on diverse projects. It generated a 19-page academic social science paper from one prompt, created games with procedurally generated art, and built a fully functional isochrone map researching 2,200+ flights and train schedules across multiple countries. Its most ambitious output was Concord, a 9.5-hour project creating software that calibrates human and AI judgments for research data analysis. Fable delegates work to cheaper models, launches multiple agents to research and verify results in parallel, and makes hundreds of independent judgment calls without user oversight. The model burns tokens at double the rate of Opus and hits security guardrails frequently. Mollick characterizes the shift: users no longer steer or do work themselves, but instead commission finished outputs from what functions as an autonomous studio.
What commenters are saying
Commenters question the author's credibility as a Wharton professor backed by corporate AI sponsors, though defenders note he cites data fairly and writes about each new model release regardless. Technical skeptics push back on capability claims, with one commenter reporting Qwen 3.7-Plus outperforms Fable for reasoning tasks. The token burn rate and 9.5-hour workflow duration divide reactions: some view long inference times as unsustainable for production use, others note that no single developer could code Concord in that timeframe. One expert flagged the design document as "slop" full of vague management-speak that obscures what the tool actually does beyond prompt-guided research and stats generation.