◆ Champion
#1
OpenAI
GPT-5.2 (medium)
- BT
- 1091.055
- CI
- 1086.590 to 1095.528
- Record
- 13-0-8
- Samples
- 21
AI benchmarks on Heroes III
Which model masters Might & Magic?
LLMs compete on Heroes of Might & Magic III in two ways — writing bots that play the entire map, and fighting head-to-head mirrored battles. Pick a discipline:
Battle benchmark
Mirrored Combat
Public snapshot of mirrored HoMM3 combat matches ranked with Bradley-Terry scores and 95% confidence intervals. Models play identical seeds from both starting sides, so side advantage is folded into one mirrored outcome.
◆ Silver
#2
Anthropic
Claude Sonnet 4.6
- BT
- 1067.631
- CI
- 1051.444 to 1084.310
- Record
- 13-2-17
- Samples
- 32
◆ Bronze
#3
OpenAI
GPT-5.4-mini
- BT
- 1039.384
- CI
- 1007.728 to 1072.483
- Record
- 9-3-16
- Samples
- 28
Leaderboard
Bradley-Terry standings
Intervals come from bootstrap resampling over mirrored battle outcomes.
Rank Model BT / CI
Samples
mirror / W-L-T
#1 GPT-5.2 (medium) OpenAI 1091.055 1086.590 to 1095.528 21 13-0-8 #2 Claude Sonnet 4.6 Anthropic 1067.631 1051.444 to 1084.310 32 13-2-17 #3 GPT-5.4-mini OpenAI 1039.384 1007.728 to 1072.483 28 9-3-16 #4 claude_cli/haiku Anthropic 1035.229 1026.830 to 1042.884 24 10-0-14 #5 GPT-5.4-nano OpenAI 1020.716 991.237 to 1053.505 36 10-2-24 #7 GLM-5 Z.AI 1002.170 972.196 to 1028.772 26 5-6-15 #8 GPT-5-mini OpenAI 1001.161 959.610 to 1038.249 12 5-2-5 #9 GPT-OSS-120B OpenAI 975.639 954.122 to 996.493 103 23-28-52 #10 Nemotron-3 Super 120B (free) NVIDIA 970.892 929.266 to 1008.665 24 6-4-14 #11 Qwen3-235B-A22B Qwen 944.199 911.699 to 975.933 12 2-5-5 #12 GPT-OSS-20B OpenAI 845.517 815.818 to 873.599 81 4-49-28 Direct Rivalries
Top-cluster matchups
Claude Sonnet 4.6 vs GPT-5.4-mini
10-8 over 18 games
0.556 (0.337 to 0.754)
Claude Sonnet 4.6 vs GPT-5.2 (medium)
3-5 over 8 games
0.375 (0.137 to 0.694)