Benchmarks · Agents & computer use · Coding · Science & research
OpenWorldEngineer (OWE-Bench)
26 long engineering projects built by domain experts, from GPU kernels to aircraft design, across science, software, ML and engineering. An agent must push each project past the best public methods; an attempt scores 0% at the untouched baseline and 100% at the task's ceiling, on held-out cases. Mean of 8 attempts per task, with a 95% interval. From Mercor.
14 models · 14 results · data as of 2026-11-12 · openworldengineer.mercor.com ↗
Leaderboard
Fable 5.2 max
More from this benchmark
2 published measuresScore vs any published measure2 measures: sub-scores, splits, counts
5 more models under All models in settings.
Breakdown
darker is better within each column · click a column to sort| Model | ||
|---|---|---|
| Fable 5.2 max | 68.3 | 2,248 |
| GPT-6 Astra max | 58.3 | 2,068 |
| Opus 5.5 max | 58.4 | 2,025 |
| GPT-6.1 Sol max | 57.1 | 2,021 |
| Sonnet 5.5 max | 52.7 | 1,975 |
| Gemini 4 Carbon high | 51.0 | 1,865 |
| Fable 5.1 max | 45.6 | 1,795 |
| Gemini 4 Argon high | 42.2 | 1,724 |
| Kimi K3.1 max | 20.6 | 1,445 |
| Muse Wonder max | 23.6 | 1,442 |
| DeepSeek V4.1 Flash max | 23.4 | 1,428 |
| GLM 5.3 max | 21.0 | 1,417 |
| Grok 4.7 xhigh | 13.8 | 1,251 |
| Inkling max | 3.8 | 1,072 |
All results
14 rows| # | Model | Effort | Score |
|---|---|---|---|
| 1 | Fable 5.2 | max | 54.6% ±8.4 |
| 2 | GPT-6 Astra | max | 44.2% ±7.7 |
| 3 | Opus 5.5 | max | 41.6% ±5.6 |
| 4 | GPT-6.1 Sol | max | 41.2% ±6.9 |
| 5 | Sonnet 5.5 | max | 39.0% ±8.2 |
| 6 | Gemini 4 Carbon | high | 32.8% ±5.6 |
| 7 | Fable 5.1 | max | 29.1% ±5.0 |
| 8 | Gemini 4 Argon | high | 24.1% ±5.8 |
| 9 | Kimi K3.1 | max | 13.6% ±3.8 |
| 10 | Muse Wonder | max | 12.7% ±4.3 |
| 11 | DeepSeek V4.1 Flash | max | 12.4% ±3.1 |
| 12 | GLM 5.3 | max | 11.8% ±3.4 |
| 13 | Grok 4.7 | xhigh | 7.0% ±2.6 |
| 14 | Inkling | max | 1.1% ±0.7 |