Code Migration v1.1 (Vals Index v2.2 subset)
Code · 2026-10-19
Vals Index v2.2's Code Migration score: the fixed subset of 50 of the 120 CLI migrations plus all 10 COBOL tasks, weighted 75% CLI and 25% COBOL, scored from the Code Migration v1.1 run (Pi harness, revised hidden tests). Vals reports that on v1.1 the subset reproduces the full result to Spearman 0.99 and 1.4 points of mean absolute error; its score is not relabeled as the full evaluation.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 78.0 |
| GPT-6.1 Sol | 75.1 |
| Gemini 4 Argon | 74.4 |
| Muse Wonder | 69.8 |
| DeepSeek V4.1 Flash 1017 | 68.5 |
| Muse Wonder Full | 61.7 |
| GLM-5.3 | 54.1 |
| DeepSeek V4 Pro 0813 | 52.0 |
| Grok 4.7 | 51.8 |
| GLM-5.3 Flash | 50.6 |
| Claude Haiku 5.5 | 50.2 |
| MiniMax M3.1 Flash | 48.2 |
| MiMo-V2.6-Pro | 47.1 |
| GPT-6 Luna | 45.5 |
| Hy4 Preview | 44.7 |
| MiMo-V2.6-Flash | 38.8 |
| Gemini 3.8 Flash | 38.2 |
| Mercury 3 | 36.9 |
| Muse Spark Open | 25.1 |
| Kimi K3 | 22.7 |
| Beam 1.0 501B A23B | 22.6 |
| Inkling-Small | 10.4 |
| Inkling | 8.6 |