RevengeBench - Halite: leaderboard

RevengeBench asks an agent to reconstruct a hidden game bot's policy from observed behavior and experiments. This series reports Halite distance reduction relative to a starter policy, in percent, under mini-SWE-agent with the best of five rounds reported. It measures how closely the recovered policy matches held-out actions; higher is better and negative values mean worse than the starter.

Metric: Distance reduction (%; best round, mini-SWE-agent). Source: revengebench.com. Saturation forecast: Not forecast. 12 models tracked. Updated 2 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.4 Mini73.3#401
2Grok 4.1 Fast66.7#429
3Kimi K2.664.4#191
4GPT-5.563.9#50
5Gemma 4 26B A4B62.5#581
6Gemma 4 31B62.3#385
7GPT-561.4#177
8GLM-5.160.8#195
9GPT-OSS-120B59.8#707
10DeepSeek V4 Pro42.6#198
11Qwen 3.6 35B A3B41.5#484
12DeepSeek V4 Flash37.1#265

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=revengebench-halite · How It Works · Data refreshed daily, snapshot 2026-11-17.