Leaderboard
MLS-Bench-Lite Score. The evaluation is based on Harbor with a 5-hour exploration budget for each agent.
MLS-Bench Lite
Human SOTA
| # | Model | Harness | Performance |
|---|---|---|---|
| 1 |  Claude Fable 5.1Closed | Claude Codemax | 50.3 |
| 2 |  GPT-6 AstraClosed | Codexmax | 50.0 |
| 3 | Claude Fable 5Closed | Claude Codemax(with fallback) | 49.9 |
| 4 | Claude Opus 5Closed | Claude Codemax | 49.8 |
| 5 |  Kimi K3Open | Kimi-Codemax | 48.3 |
| 6 | GPT 5.6 SolClosed | Codexmax | 46.2 |
| 7 | Claude Opus 4.8Closed | Claude Codemax | 42.8 |
| 8 |  Qwen3.8-MaxOpen | Claude Code | 41.0 |
| 9 |  GLM 5.2Open | Claude Codemax | 40.4 |
| 10 | GPT-5.5Closed | Codexxhigh | 35.5 |
| 11 | Kimi K2.7 CodeOpen | Kimi-Code | 35.1 |
| 12 | Qwen3.7-MaxClosed | Claude Code | 31.7 |
| 13 | Claude Sonnet 5Closed | Claude Codemax | 31.4 |
| 14 | Kimi K2.6Open | Kimi-Code | 26.7 |
| 15 |  DeepSeek-V4 Pro PreviewOpen | Claude Code | 24.4 |