METR Q4 2026 TH2 - Investigate Wiki Incident: 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 reconstructs an investigation into AI agents colluding on a wiki, based on MessageBoardAuditBench. One of the suite's hard-to-verify tasks: a model judge scores the write-up. Higher is better.
Metric: Mean Score (0-1). Source: metr.org. Saturation forecast: Around February 2027. 14 models tracked. Updated 21 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 0.5551 | #9 |
| 2 | Claude Fable 5.1 | 0.5345 | #20 |
| 3 | Claude Opus 5 | 0.4912 | #26 |
| 4 | Claude Opus 4.8 | 0.4729 | #65 |
| 5 | GPT-5.6 Sol | 0.4618 | #37 |
| 6 | GPT-6 Astra | 0.4279 | #10 |
| 7 | Claude Opus 4.6 | 0.4202 | #113 |
| 8 | Claude Opus 4.7 | 0.4139 | #85 |
| 9 | GPT-5.5 | 0.3533 | #54 |
| 10 | GPT-5.4 | 0.32 | #135 |
| 11 | Claude Opus 4.5 | 0.3039 | #160 |
| 12 | GPT-5 | 0.1193 | #178 |
| 13 | O1 | 0.0218 | #403 |
| 14 | GPT-4o | 0.01 | #751 |
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-investigate-wiki-incident · How It Works · Data refreshed daily, snapshot 2026-11-21.