METR: AI R&D time horizons double every 3.6 months on its combined TH1.1 and TH2 suite
METR's Q4 2026 update covers public models up to Opus 5.5 and GPT-6 Astra and says public evidence cannot rule out drastic AI R&D acceleration in the next couple of years.
What happened
METR is a group that tests what AI models can do. It reports that AI models are getting better fast at coding and optimization work. On its tasks, models usually match or beat skilled humans. The group says its tests are also getting harder to trust.
Why it matters to you
If AI can speed up AI research itself, progress could outrun the ability of people to understand and respond to risks. METR says current public evidence cannot rule that out within a couple of years.
The takeaway
METR says AI is improving quickly at technical work and that measuring it well is getting harder.
What it means for you
Agent swarms looked helpful on verifiable tasks with a clear metric and fast feedback compared to best-of-k, and had unclear or negative effects on other tasks.
METR says models perform worse on tasks that are costlier to turn into RL environments, but they are improving at similar rates on cheaper and costlier tasks.