MTGBench - Native Reasoning - Russian - Overall - Win Rate: leaderboard
MTGBench Native Reasoning protocol in Russian, Overall: win rate among completed episodes. Seven multi-turn puzzle games measure deduction, planning, information gathering and consistency. Incomplete/missing episodes follow the native exclusion policy; completed sample sizes and exact prompt/native-reasoning modes are retained. Overall equally averages seven game means.
Metric: Win Rate among completed episodes (%). Source: github.com. Saturation forecast: Rough model projection: around 2026. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | gemini-3.1-pro-preview (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 96.57 | |
| 2 | gemini-3-flash-preview (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 94.14 | |
| 3 | gpt-5.4 (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 93.43 | |
| 4 | gemini-3.1-flash-lite (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 84.22 | |
| 5 | qwen3.5-35b-a3b (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 81.32 | |
| 6 | deepseek-v3.1-terminus (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 80.71 | |
| 7 | deepseek-v4-pro (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 80.43 | |
| 8 | qwen3-235b-a22b-thinking-2507 (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 79 | |
| 9 | gpt-oss-120b (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 77.84 | |
| 10 | gpt-5.4-mini (MTGBench Run: mode=reasoning;effort=unreported;release=ffe4adc59146) | 76.14 |
Interactive version: theaggregate.ai/benchmark?slug=mtgbench-native-reasoning-russian-overall-win-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.