AHK-Eval
Code · 2026-06-11
AHK-Eval is a HumanEval-style AutoHotkey v2 benchmark of 36 independent function-generation tasks. Each submission is parsed and executed on the pinned AutoHotkey v2.1-alpha.30+Console fork against 181 hidden functional cases. A task passes only when every hidden case passes; the canonical cold protocol uses one fresh API call per task, no tools or retries, temperature 0.2, and an 8,000-token output cap.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 100.0 |
| Opus 4.8 | 100.0 |
| GPT-5.6 Sol Pro | 97.2 |
| GPT-5.5 | 97.2 |
| Gemini 3.1 Pro Preview | 97.2 |
| Sonnet 4.6 | 97.2 |
| GPT-5.6 Sol | 94.4 |
| GPT-5.6 Luna Pro | 94.4 |
| Kimi K3 | 94.4 |
| GPT-5.6 Luna | 91.7 |
| GPT-5.6 Terra | 88.9 |
| Kimi K2.6 | 88.9 |
| GPT-5.6 Terra Pro | 86.1 |
| Grok 4.5 | 86.1 |
| Grok 4.3 | 86.1 |
| Muse Spark 1.1 | 83.3 |
| MiniMax M3 | 83.3 |
| GLM-5 | 80.6 |
| Aion 3.0 | 75.0 |
| DeepSeek-V4-Pro | 75.0 |
| GPT-5.1 | 69.4 |
| GLM-5.2 | 66.7 |
| Aion 3.0 Mini | 61.1 |
| Hy3 | 61.1 |
| Qwen3-Coder-480B-A35B-Instruct | 50.0 |
| Mistral Large 3 675B Instruct 2512 | 41.7 |
| Laguna XS 2.1 | 25.0 |