METR's new TH2 eval: capability doubling time shrinks to 1.65 months, could saturate by July 2027
josogood · reddit · 2026-11-21
METR's November 2026 update introduces TH2, a robust successor to its saturated time-horizon benchmark, measuring judgment and hard-to-verify tasks for models released through Sept 2026. Key findings: verified-task doubling time tightened from 3.07 to 1.65 months; parts of TH2 could saturate by July 2027. Models excel at coding/math but lag humans on costly tasks and judgment without empirical feedback; tested coverage tops out at Opus 5.5 / GPT-6 Astra.
Related event: METR Update: AI R&D Capabilities Rising Fast and Getting Harder to Measure(5 posts)→
More from AGI Musings
- RSI is outpacing AI 2027 forecasts: the final 18 months of white-collar work — ChrisGPT · 2026-11-21
- Titanic violin meme sweeps AI circle: 'a privilege escaping the permanent underclass with you all' — D3VAUX · 2026-11-21
- Why does Yudkowsky's 2001 text sound exactly like LLM 'Claude-slop'? — paulnovosad · 2026-11-21
- Compute overbuild skeptic recants: efficiency gains cut AI compute cost ~4x per year — teortaxesTex · 2026-11-21
- The classic waitbutwhy intelligence-explosion chart resurfaces as models keep jumping — burny_tech · 2026-11-21
- Vibe-coding culture: everyone builds over the holidays, few realize when to log off — haydendevs · 2026-11-21