MARS-Bench - Context Retrieval: 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 retrieve events from earlier turns given a time anchor or interval; 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)87.06#271
2GPT-4.5 (Preview)66.65#499
3O1 (2024-12-17)64.48#314
4Claude 3.7 Sonnet59.77#476
5Doubao-1.5-Thinking-Pro55.64#257
6DeepSeek R149.23#487
7Doubao-1.5-Pro-32k46.81#696
8DeepSeek V3 (1226)37.08#807
9Qwen 2.5 Max31.9#966
10GPT-4o (2024-11-20)31.26#864
11Qwen 2.5 72B Instruct30.41#1119
12GPT-4.1 Mini30.17#793
13GLM-Z1-Air22.49

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-context-retrieval · How It Works · Data refreshed daily, snapshot 2026-11-17.