Ko-AgentBench - L5 Error Handling & Robustness: leaderboard
Ko-AgentBench level 5 — Korean-language agent tasks testing recovery from tool errors and malformed or incomplete results; accuracy.
Metric: Adaptive Routing Score (%). Source: huggingface.co. Saturation forecast: Not forecast. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.1 Fast | 34.75 | #217 |
| 2 | GPT-5 | 30 | #98 |
| 3 | Nova 2.0 Lite (Non-reasoning) | 26.25 | #417 (Nova 2.0 Lite) |
| 4 | GLM-4.6V | 26.01 | #316 |
| 5 | GPT-4o | 25.83 | #343 |
| 6 | Qwen 3 Next 80B A3B | 25.42 | #320 |
| 7 | Claude Haiku 4.5 | 22.83 | #280 |
| 8 | GPT-4o Mini | 21.71 | #599 |
| 9 | DeepSeek V3.1 | 21.33 | #264 |
| 10 | Claude Sonnet 4.5 | 18.92 | #142 |
| 11 | Nova Lite | 13.76 | #754 |
| 12 | Gemini 2.5 Pro | 12.5 | #152 |
| 13 | Gemini 2.5 Flash Lite | 12.5 | #422 |
| 14 | Gemini 2.5 Flash | 10 | #245 |
| 15 | GPT-5 Mini | 9.17 | #186 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ko-agentbench-l5-error-handling-robustness · How It Works · Data refreshed daily, snapshot 2026-10-20.