METR Time Horizons 2.0 returns: 3.6-month doubling time, GPT-6 Astra hits 40.3 hours
scaling01 · x · 2026-11-21
scaling01 highlights METR's Time Horizons 2.0 update (back after TH1.1 data went dark):
- Doubling time: 3.6 months, close to OpenAI's own 3.8-month estimate; doubling on TH2 (verifiable) tasks is notably faster than on the old TH1.1
- Latest scores: GPT-6 Astra at 40.3 hours, Opus 5.5 at 39.0 hours
- Huge uncertainty: the 95% CI spans roughly 11.9 hours to 110.2 days — nearly two orders of magnitude
- The chart shows bootstrap doubling-time estimates of 3.07 mo (TH1.1), 1.65 mo (TH2), 1.79 mo (MirrorCode)
More from Models
- Riemann proof was human-AI collaboration from Anthropic — and not a swarm — BlackHC · 2026-11-21
- Google Search AI leaks its chain-of-thought and internal formatting tags to users — ConsistentMap5775 · 2026-11-21
- SenseNova Loop 9B on a single RTX 4080: 256k context for 16GB VRAM GPUs — roofkid · 2026-11-21
- Google's subscription page leaks unannounced "Hydrogen" model in AI Plus tier — giffmana · 2026-11-21
- xAI staff responds to Grok bot computer failures, escalates support ticket — mattyp · 2026-11-21
- Looped transformers: more looping boosts reasoning but with diminishing returns past 3 passes and fewer CoT tokens — nrehiew_ · 2026-11-21