RAMP: leaderboard
Long-horizon agentic software engineering: an agent on a shared OpenHands SDK backend completes six serially dependent stages of a C compiler from the YatCC project (environment setup, lexer, parser, LLVM IR generation, IR optimization, RV64 assembly), with failed stages not restored. Mean Reward is a completion-weighted average of the stage test scores (0-100, higher is better), and the untouched scaffold alone scores 23.38 (Ouyang et al., Sun Yat-sen University, 2026).
Metric: Mean Reward (0-100, weighted task score). Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 93.39 | #75 |
| 2 | DeepSeek V4 Pro | 85.34 | #181 |
| 3 | DeepSeek V4 Flash | 66.48 | #254 |
| 4 | GPT-5.5 | 65.91 | #48 |
| 5 | Qwen 3.6 Max Preview | 56.72 | #155 |
| 6 | Kimi K2.5 | 40.24 | #252 |
| 7 | Kimi K2.6 | 33.94 | #173 |
| 8 | GLM-4.6 | 30.88 | #485 |
| 9 | Qwen 3.5 Plus | 29.63 | #227 |
| 10 | GLM-4.7 | 29.5 | #365 |
| 11 | GLM-5.1 | 27.26 | #175 |
| 12 | Qwen 3 Coder Flash | 25.97 | #1624 |
| 13 | MiniMax-M2.5 | 25.09 | #590 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ramp · How It Works · Data refreshed daily, snapshot 2026-11-12.