MARS-Bench: leaderboard

Each model follows an ultra multi-turn chat over text play-by-play of 120 games (60 English NBA, 40 Chinese NBA, 20 English NHL) and answers questions inserted between turns; the score is the mean of four category scores of 30 games each. A DeepSeek-V3-0324 judge awards checklist points built from the play-by-play. Unlike MT-Bench's short dialogues, the questions depend on facts many turns back.

Metric: Score (0-100). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro (Preview 05-06)72.44#271
2O1 (2024-12-17)59.62#314
3GPT-4.5 (Preview)53.33#499
4Doubao-1.5-Thinking-Pro52.62#257
5DeepSeek R145.42#487
6Claude 3.7 Sonnet41.21#476
7Doubao-1.5-Pro-32k38.88#696
8GPT-4o (2024-11-20)35.83#864
9DeepSeek V3 (1226)33.16#807
10GPT-4.1 Mini31.23#793
11Qwen 2.5 Max30.41#966
12Qwen 2.5 72B Instruct29.21#1119
13GLM-Z1-Air25.84

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

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