MARS-Bench - Task Switching: leaderboard

Each model, in an ultra multi-turn chat over the text play-by-play of 30 games (15 English NBA, 10 Chinese NBA, 5 English NHL), must answer the match questions interleaved with unrelated math questions (the score covers the match questions only; the paper reports the math separately); the score is the category score. A DeepSeek-V3-0324 judge awards checklist points built from the play-by-play.

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)66.72#271
2O1 (2024-12-17)58.28#314
3Doubao-1.5-Thinking-Pro50.69#257
4GPT-4.5 (Preview)40.74#499
5DeepSeek R139.4#487
6GPT-4o (2024-11-20)36.12#864
7Claude 3.7 Sonnet34.15#476
8Doubao-1.5-Pro-32k32.28#696
9DeepSeek V3 (1226)29.63#807
10GPT-4.1 Mini28.13#793
11Qwen 2.5 72B Instruct23.97#1119
12Qwen 2.5 Max23.22#966
13GLM-Z1-Air20.76

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-task-switching · How It Works · Data refreshed daily, snapshot 2026-11-17.