Heidelberg AI News

Video · Wes Roth ·

Anthropic BROKE Fable 5.2

In short

The video examines Anthropic’s claimed Fable 5.2 and Mythos 5.2 release, focusing on their reported strengths in long-horizon coding, robotics, hardware design, and physics, alongside deliberate training limits in accelerator-kernel optimization and LLM architecture research. It argues these models may correspond to METR’s Model 3 and could signal progress toward autonomous AI research, illustrated through a multi-agent operating-system effort and a game benchmark.

Key points

  • Fable 5.2 and Mythos 5.2 are presented as exceptionally capable models for complex, long-horizon tasks.
  • Anthropic reportedly capped performance near Opus 5.5 in accelerator-kernel optimization and LLM architecture research during training.
  • The video identifies the release with an early METR Model 3 and discusses a possible unreleased Model 4.
  • A secret agent scaffold, compared with NanoGPT and AutoResearcher, may enable large-scale autonomous experiments.
  • More than 1,000 agents reportedly worked for 8 days to produce a functional operating-system prototype.
  • Fable 5.2 rapidly improved a 2D game’s graphics and presentation in a live benchmark.
  • Higher-tier plans reportedly receive up to 3 usage resets with the release.

Written by an AI model from the video’s captions and description.

Watch on YouTube The opening of the article at youtube.com.

In this story

A tool or model opens in the Heidelberg AI Curriculum: what it does, what it costs, how to start. A repo opens on GitHub.

Related stories

All stories