METR Q4 2026 TH2 - Budget NanoGPT (Expenditure Horizon): leaderboard

METR's Q4 2026 capability update (20 November 2026): what humans would have to spend to match the agent on the NanoGPT speedrun. METR prices human speedrun records at $2,500 per 1% of remaining training time cut and reads off the spend that reaches the agent's mean best speedup (agents themselves spent about $1,000 to $2,000 in tokens and GPUs). Human costs come from contributor interviews, so the figures are approximate. US dollars; higher is better.

Metric: Expenditure Horizon (USD). Source: metr.org. Saturation forecast: Around March 2028. 6 models tracked. Updated 21 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-6 Astra25112#10
2Claude Fable 5.121402#20
3Claude Opus 5.520099#9
4Claude Opus 510828#26
5Claude Fable 52873#33
6GPT-5.6 Sol2865#37

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=metr-q4-2026-th2-budget-nanogpt-expenditure-horizon · How It Works · Data refreshed daily, snapshot 2026-11-21.