BrowseComp Extended (Mercor held-out)
Search
Mercor's held-out Extended twin of OpenAI's BrowseComp: 100 Mercor questions that require agents to persistently search the web for hard-to-find, entangled facts, scored by answer accuracy. The public twin is BrowseComp (browsecomp); Mercor's rerun of its 130-question subset is browsecomp-mercor-130. Mercor does not publish the held-out tasks or say how they were drawn or written, so this is its own benchmark, not a component of the public one.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Argon | 72.0 |
| GPT-6.1 Sol | 67.7 |
| GPT-6 Sol | 63.2 |
| Claude Opus 5.5 | 60.5 |
| GPT-6 Luna | 47.2 |
| Claude Fable 5 | 44.5 |
| Claude Opus 5 | 44.0 |
| Gemini 3.8 Flash | 41.5 |
| Grok 4.7 | 39.5 |
| Muse Spark 1.2 | 32.5 |
| GPT-5.6 Sol | 32.3 |
| Opus 4.8 | 30.0 |
| Kimi K3 | 28.5 |
| Claude Haiku 5.5 | 27.5 |
| DeepSeek V4 Pro 0813 | 27.0 |
| Gemini 3.7 Flash | 25.5 |
| GPT-5.6 Terra | 25.3 |
| Beam 1.0 501B A23B | 25.0 |
| GLM-5.2 | 23.0 |
| Grok 4.6 | 21.5 |
| Sonnet 5 | 19.0 |
| GPT-5.5 | 19.0 |
| GPT-5.6 Luna | 16.5 |
| Opus 4.7 | 14.2 |
| Sonnet 4.6 | 14.2 |
| Grok 4.5 | 14.2 |
| GPT-5.4 | 14.0 |
| Gemini 3.6 Flash | 13.8 |
| Gemini 3.1 Pro Preview | 10.7 |
| Kimi K2.7 Code | 8.0 |
| MiniMax M3 | 7.7 |
| Inkling | 4.5 |
| DeepSeek-V3.2 | 3.3 |
| Gemini 3.5 Flash | 0.7 |
| MiniMax M2.7 | 0.5 |