KOPA-Bench - Action Accuracy: leaderboard
KOPA-Bench evaluates multi-step tool use over Korean public APIs on 145 tasks, with four rollouts per task. Pass@1 averages single-attempt success, pass@4 measures at least one success, and Action measures tool-call accuracy against golden actions. The paper's GRPO-tuned models remain distinct from their base checkpoints. This series reports KOPA-Bench - Action Accuracy from 2609.05395v1, S3.T1: Tool-call Accuracy (%). Higher is better.
Metric: Tool-call Accuracy (%). Source: arxiv.org. Saturation forecast: Around July 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 58.81 | #224 |
| 2 | Gemma 4 26B A4B (IT) | 50.61 | #573 |
| 3 | Claude Sonnet 4.6 | 42.69 | #159 |
| 4 | Qwen 3.5 27B | 42.13 | #368 |
| 5 | gemma-4-E4B-it | 40.69 | #1548 |
| 6 | Ministral-3-8B-Instruct-2512 | 37.97 | #1948 |
| 7 | Qwen 3.5 9B | 34.07 | #862 |
| 8 | EXAONE 4.5 33B | 26.7 | #772 |
| 9 | Qwen 3.5 4B | 21.4 | #1255 |
No result here: #4 Claude Opus 5.5, #5 GPT-6 Astra, #9 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kopa-bench-action-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-08.