Leaked METR docs: Claude Mythos 5.2 is Model 3 Early, may fully automate Anthropic's AI R&D

davidmanheim · x · 2026-11-10

A retweet of "Mythos 5.2 is Model 3" with screenshots of Anthropic/METR documents — the substance is in the images:

METR's pre-deployment testing notes for Claude Mythos 5.2 identify its snapshot as Model 3 Early in METR's independent assessment
METR concludes Mythos 5.2 is a substantial step above Fable 5.1 on verifiable tasks (notably strong Budget NanoGP Speedrun optimization) and is uncertain but thinks it can fully automate Anthropic's AI R&D process
Evidence cited: solved a clear majority of an internal month-long research benchmark, nearly solved most internal long-horizon tasks in half to three-quarters the time, completed several multi-day fully autonomous software builds with self-verification
METR chart: as of Oct 2026, benchmark scores (including internal models) are growing at 1.9× the pre-Oct-2025 rate (+15/yr → +29/yr, 95% CI 1.5–2.8)

If authentic, these screenshots are significant first-hand material on Anthropic's internal model lineage and AI R&D automation progress.

More from AGI Musings