DuelLabBenchmark
Main site
Benchmark menu

Rankings

Current leaderboard score across public games (0–100)

Showing top 24 of 34 benchmarked models (updates when chart loads)

Scale: relative 0-100

  1. 71.8
    Claude Fable 5
  2. 55.1
    GPT-5.4
  3. 54.0
    GPT-5.6 Sol
  4. 51.1
    Claude Opus 4.8
  5. 49.8
    O3
  6. 48.9
    GPT-5.5
  7. 44.4
    Gemini 3.5 Flash
  8. 42.0
    GPT 5
  9. 41.9
    Claude Opus 4.5
  10. 40.9
    GPT-5.6 Terra
  11. 39.8
    DeepSeek V4 Pro
  12. 39.6
    Claude Sonnet 5
  13. 39.2
    Kimi K2.7 Code
  14. 36.8
    MiMo-V2.5-Pro
  15. 36.6
    GLM-5.2
  16. 36.6
    GPT 4.1
  17. 35.3
    GPT-5.4 Mini
  18. 34.9
    Qwen3.7 Max
  19. 34.0
    GPT-5.4 Nano
  20. 32.0
    GPT-5.6 Luna
  21. 31.7
    Grok Build 0.1
  22. 29.2
    Minimax M3
  23. 29.0
    Qwen3.7 Plus
  24. 28.5
    Hy3 Preview
View

Models

Overall leaderboard

View: Overall Reasoning levels: 3 Per-game: Filterable
Overall Models leaderboard for DuelLab Benchmark
Rank Model Score Levels Entries Reasoning spread Signal
1Claude Fable 571.821573.7 / 69.8 / —87.5u 12.5
2GPT-5.455.132466.8 / 49.1 / 49.489.8u 10.2
3GPT-5.6 Sol54.032472.4 / 52.3 / 37.289.3u 10.7
4Claude Opus 4.851.132268.6 / 41.5 / 43.387.5u 12.5
5O349.821057.0 / 42.7 / —100.0u 0.0
6GPT-5.548.932267.6 / 47.2 / 31.892.6u 7.4
7Gemini 3.5 Flash44.421438.9 / 49.8 / —87.5u 12.5
8GPT 542.021543.8 / 40.2 / —92.5u 7.5
9Claude Opus 4.541.932342.4 / 37.2 / 45.986.9u 13.1
10GPT-5.6 Terra40.932451.9 / 37.9 / 33.189.1u 10.9
11DeepSeek V4 Pro39.821342.3 / — / 37.490.7u 9.3
12Claude Sonnet 539.632344.0 / 36.5 / 38.287.8u 12.2
13Kimi K2.7 Code39.216— / 39.2 / —100.0u 0.0
14MiMo-V2.5-Pro36.8212— / 40.5 / 33.193.2u 6.8
15GLM-5.236.636236.9 / 36.6 / 36.382.7u 17.3
16GPT 4.136.618— / — / 36.689.8u 10.2
17GPT-5.4 Mini35.335947.6 / 32.6 / 25.678.4u 21.6
18Qwen3.7 Max34.9214— / 36.1 / 33.786.5u 13.5
19GPT-5.4 Nano34.036545.6 / 33.6 / 23.084.7u 15.3
20GPT-5.6 Luna32.032455.0 / 29.3 / 11.787.5u 12.5
21Grok Build 0.131.718— / 31.7 / —92.4u 7.6
22Minimax M329.2213— / 19.7 / 38.787.5u 12.5
23Qwen3.7 Plus29.0212— / 39.6 / 18.4100.0u 0.0
24Hy3 Preview28.521328.6 / — / 28.387.5u 12.5
25MiMo-V2.527.733722.1 / 24.5 / 36.587.5u 12.5
26Nemotron 3 Ultra 550B A55B25.832118.9 / 26.9 / 31.587.5u 12.5
27Gemma 4 31B24.6210— / 28.3 / 20.8100.0u 0.0
28Step 3.7 Flash24.12925.9 / 22.3 / —100.0u 0.0
29DeepSeek V4 Flash23.534428.0 / 21.0 / 21.685.1u 14.9
30GPT-OSS 120B23.521227.7 / 19.3 / —87.5u 12.5
31Gemini 3.1 Flash Lite17.031621.7 / 19.3 / 10.1100.0u 0.0
32Nex N2 Pro26.1212— / 20.7 / 31.687.5u 12.5
33North Mini Code22.7211— / 27.4 / 18.090.7u 9.3
34Mistral Medium 3.515.32815.1 / — / 15.5100.0u 0.0

The overall view averages performance across available reasoning variants. Use the controls above to filter models or rerank both the bars and table for one game; models without a score for the selected game follow the scored rows.

Scoring method

How this is scored

1. Generate players

Each model is asked to create a program that can play every benchmark game.

2. Play matches

The generated programs compete head-to-head, with both players receiving comparable opportunities.

3. Score each game

Results and how certain they are produce a score from 0 to 100 for each game.

4. Combine games

Game scores are combined into the main leaderboard score. A known failure to create a usable player contributes zero.

5. Keep settings clear

The reasoning settings are combined into one row for each model.

6. Add context

Detailed tables provide uncertainty, match records, and other clues for careful comparison.