What it feels like to work with Mythos

Open original

The author tests Anthropic’s Mythos-class model, Claude 5 Fable, on tasks from epic poems to complex isochrone maps and research calibration software. Fable autonomously delegates work to cheaper agents, executes multi‐hour workflows, and produces sophisticated outputs, but its decision process remains a black box, shifting the user’s role from hands‐on builder to outcome judge.

  • Claude 5 Fable autonomously delegated research to dozens of Sonnet agents (and spun up adversarial agents for edge cases like Pitcairn Island and Grise Fjord) to build a polished isochrone map in hours.
  • It ran a nine-and-a-half-hour autonomous build of "Concord," a research calibration tool, complete with a 19-page design spec and production-ready code.
  • The user's role shrank to prompting and light feedback while the model made hundreds of unseen micro-decisions, turning oversight into outcome judging rather than hands-on building.

Claude 5 Fable, the first public Mythos-class model, blew past every AI I’ve tried. In side-by-side tests it outperformed other public models by a wide margin and stuck with multi-page specs for hours. It whipped up a sophisticated social-science paper from a single prompt and one piece of feedback, then turned around a ten-page, all-“s” epic rhyming poem. For a more practical demo, I had it spin up several games using only code—no images—like a coin-flip strategy game, a self-aware snake romp, and a descent-into-darkness adventure. Each one began with a vague instruction, followed by light “make it better” tweaks, yet emerged fully playable. The real kicker arrived when I asked Fable to build an isochrone map showing how far you can travel from a city in a set time. Previous models floundered at the research and judgment calls. Fable launched dozens of Claude Sonnet agents to scour schedules—2,200 flights, TGV and Shinkansen timetables, road-speed studies—and then started coding. When I flagged gaps (remote spots like Greenland), it spun up adversarial research agents to nail down ship routes to Pitcairn Island and flight connections to Grise Fjord. In under a few hours it delivered a polished, 1881-style map with customizable cities, visual tweaks, full data sources and methods listed. The most ambitious test was a nine-and-a-half-hour run to build “Concord,” a tool for calibrating human and AI judgments on messy research data. Fable wrote a 19-page design spec, generated production-ready code, and handled dataset integration, calibration algorithms and analytical reporting. I identified a handful of expert-level tweaks, asked for fixes, and it updated the system. The result is a codebase researchers can fork or refine—something the community long needed but never funded. At each step, I played a minimal role: issue a prompt, give light feedback, and the model carried out hundreds of micro-decisions I neither saw nor controlled.

Ask about this article