Ko-AgentBench - L1 Single Tool Call: leaderboard
Ko-AgentBench level 1 — Korean-language agent tasks requiring one correctly selected and parameterized tool call; accuracy.
Metric: Argument Accuracy (%). Source: huggingface.co. Saturation forecast: Not forecast. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nova 2.0 Lite (Non-reasoning) | 79.55 | #417 (Nova 2.0 Lite) |
| 2 | Qwen 3 Next 80B A3B | 77.27 | #320 |
| 3 | Grok 4.1 Fast | 75 | #217 |
| 4 | GPT-5 | 72.73 | #98 |
| 5 | Gemini 2.5 Pro | 70.45 | #152 |
| 6 | GPT-5 Mini | 70.45 | #186 |
| 7 | GLM-4.6V | 68.18 | #316 |
| 8 | Claude Sonnet 4.5 | 65.91 | #142 |
| 9 | GPT-4o Mini | 65.91 | #599 |
| 10 | Gemini 2.5 Flash Lite | 63.64 | #422 |
| 11 | Gemini 2.5 Flash | 61.36 | #245 |
| 12 | Claude Haiku 4.5 | 61.36 | #280 |
| 13 | GPT-4o | 59.09 | #343 |
| 14 | Nova Lite | 59.09 | #754 |
| 15 | DeepSeek V3.1 | 54.55 | #264 |
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-l1-single-tool-call · How It Works · Data refreshed daily, snapshot 2026-10-20.