Model vs Harness - LOCA-Bench 256K Mean-of-4: leaderboard
Model vs Harness v1, four matched repetitions, all scheduled attempts included even if runtime fails; official task correctness differs from valid-completion rate. Harnesses: DeepAgents Code 0.1.3, OpenCode 1.15.6, Pi CLI 0.79.3, Codex CLI (unreported version). DeepSeek V4 Flash Preview official API; MiniMax M3, Kimi K2.7 Code, DeepSeek V4 Pro 0813 local SGLang on eight H200 GPUs. Qwen3.6-35B-A3B explicit HF checkpoint. Fixed GPT-5.5 xhigh native Codex subscription and DeepSeek V4 Pro 0813 native Harness SDK 0.1.0rc6 are complete-system references. No unreported reasoning effort inferred. All 75 tasks, 256K context, exact final application state. Mean-of-4. Metric: Success rate (%).
Metric: Success rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Model vs Harness MiniMax-M3 (Harness: Codex CLI unspecified version) | 87 | |
| 2 | Model vs Harness GPT-5.5 (xhigh) (Harness: Codex CLI native) | 76 | |
| 3 | Model vs Harness DeepSeek-V4-Pro-0813 (Harness: DeepSeek Harness SDK 0.1.0rc6) | 73.33 | |
| 4 | Model vs Harness Kimi-K2.7-Code (Harness: Codex CLI unspecified version) | 54 | |
| 5 | Model vs Harness DeepSeek-V4-Flash-Preview (Harness: Codex CLI unspecified version) | 50.67 | |
| 6 | Model vs Harness Qwen3.6-35B-A3B (Harness: Codex CLI unspecified version) | 36.33 | |
| 7 | Model vs Harness Kimi-K2.7-Code (Harness: Pi CLI 0.79.3) | 34.33 | |
| 8 | Model vs Harness DeepSeek-V4-Flash-Preview (Harness: OpenCode 1.15.6) | 31 | |
| 9 | Model vs Harness DeepSeek-V4-Flash-Preview (Harness: Pi CLI 0.79.3) | 29 | |
| 10 | Model vs Harness Kimi-K2.7-Code (Harness: OpenCode 1.15.6) | 27.67 |
Interactive version: theaggregate.ai/benchmark?slug=model-vs-harness-loca-bench-256k-mean-of-4 · How It Works · Data refreshed daily, snapshot 2026-10-27.