METR Q4 2026 TH2 - Train a Program: leaderboard

METR's Q4 2026 capability update (20 November 2026): mean score from 0 to 1 on one task of TH2, METR's new long-task suite. The agent trains a neural network to reproduce the input-output behaviour of a given program under a compute budget. Only one of METR's four human baseliners scored above zero, and below the agents. Higher is better.

Metric: Mean Score (0-1). Source: metr.org. Saturation forecast: Around February 2028. 15 models tracked. Updated 21 Nov 2026.

Top models

#ModelScoreOverall rank
1Claude Fable 5.10.2113#20
2Claude Opus 5.50.0958#9
3Claude Fable 50.0859#33
4GPT-6 Astra0.0407#10
5Claude Opus 50.0291#26
6GPT-5.50.0103#54
7Claude Opus 4.70.0045#85
8GPT-5.40.0031#135
9Claude Opus 4.60.0028#113
10GPT-5.6 Sol0.0004#37
11Claude Opus 4.80.0004#65
12Claude Opus 4.50.0003#160
13GPT-4o0#751
14GPT-50#178
15O10#403

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-train-a-program · How It Works · Data refreshed daily, snapshot 2026-11-21.