Summary
Wes Roth walks through the release of Anthropic’s Fable 5.2 and Mythos 5.2, which he describes as the strongest models available to the public. He highlights long-horizon tasks where the models reportedly outperform Opus 5.5, including robotic control, hardware engineering, and a theoretical physics result involving black holes that was checked by a physicist and cross-verified by a second Claude computation.
The video then turns to the recent METR report and the sharp jump in capability around the Mythos release. Roth explains his reading of Anthropic’s internal model tiers, arguing that the publicly released Fable and Mythos 5.2 correspond to an earlier checkpoint than the company’s more advanced internal models. He also discusses claims that Anthropic deliberately limited certain capabilities during training, and speculates about the redacted details of what is accelerating AI research.
Additional segments cover an experiment in which over 1,000 agents worked for eight days to build a prototype operating system from an empty folder, Anthropic’s apparent anti-distillation measures around reasoning traces, and a comparison to Andrej Karpathy’s nanoGPT and autoresearch projects as a possible model for the hidden agent scaffold. The discussion touches on recursive self-improvement and what faster AI-driven research could mean for future model releases.
📺 Source: Wes Roth · Published November 10, 2026
🏷️ Format: News Analysis
