K-BrowseComp - Synthetic: leaderboard
The synthetic split of K-BrowseComp (2026): Pass@1 accuracy in percent on 100 agent-generated Korean browsing questions that were kept only when GPT-5.4-mini and Gemini 3 Flash failed them, under the same 10-search agent harness. Higher is better.
Metric: Pass@1 accuracy (%) on the 100-question synthetic split (agent-generated Korean browsing questions kept only when gpt-5.4-mini and gemini-3-flash-preview failed them); same harness and grading. Source: arxiv.org. Saturation forecast: Around July 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 26 | #49 |
| 2 | DeepSeek V4 Pro | 22 | #188 |
| 3 | GLM-5.1 | 19 | #179 |
| 4 | Gemma 4 31B (IT) | 17 | #342 |
| 5 | Qwen 3.6 35B A3B | 15 | #462 |
| 6 | K-EXAONE | 13 | #708 |
| 7 | Gemini 3.1 Flash Lite | 11 | #364 |
| 8 | HyperCLOVA X SEED Think (32B) | 2 | #1437 |
| 9 | A.X 4.0 | 1 | #863 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=k-browsecomp-synthetic · How It Works · Data refreshed daily, snapshot 2026-11-13.