DSGym - DSPredict Hard - Above Median: leaderboard
DSPredict-Hard, DSGym's split of recent, still-open Kaggle competitions with complex data and objectives, scored on the private leaderboard and run by DSGym's default CodeAct agent (Nie et al., 2026) in an isolated Jupyter container at temperature 0 with all tools disabled. Share of competitions whose submission beats the leaderboard median, in percent; higher is better.
Metric: Above Median Rate (%; private leaderboard). Source: arxiv.org. Saturation forecast: Around March 2028. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 (High) | 14.3 | #225 (GPT-5.1) |
| 2 | GPT-5.1 (Non-reasoning) | 10.3 | #225 (GPT-5.1) |
| 3 | DeepSeek V3.1 | 7.1 | #500 |
| 4 | GPT-5.1 (Medium) | 7.1 | #225 (GPT-5.1) |
| 5 | Claude Sonnet 4.5 | 4.8 | #244 |
| 6 | Claude Sonnet 4 | 4.8 | #397 |
| 7 | GPT-5 (Medium) | 2.4 | #163 (GPT-5) |
| 8 | Qwen 2.5 7B Instruct | 0 | #2583 |
| 9 | Kimi K2 0905 | 0 | #555 |
| 10 | Qwen 3 4B Instruct | 0 | #1660 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dsgym-dspredict-hard-above-median · How It Works · Data refreshed daily, snapshot 2026-11-12.