Calibration
React Native Evals → Vibe Code Bench
Vibe Code Bench is estimated from React Native Evals with a offset logistic curve fitted on 7 models measured on both: y = 0.2767 + (0.6655 − 0.2767) / (1 + exp(−166.46·(x − 0.8029))), R² = 0.94, cross-validated error 6.5 pp. It is used for 9 estimates.
| Estimated model | React Native Evals | Vibe Code Bench | Source |
|---|---|---|---|
| Composer 2 | 96.1% | 66.5% | estimated ± 6.5 pp, low confidence |
| Composer 2 Fast | 94.9% | 66.5% | estimated ± 6.5 pp, low confidence |
| DeepSeek V3.2 | 71.5% | 27.7% | estimated ± 6.5 pp, low confidence |
| Gemma 4 31B | 75.2% | 27.7% | estimated ± 6.5 pp, low confidence |
| GLM-5 | 74.8% | 27.7% | estimated ± 6.5 pp, low confidence |
| GPT-OSS 120B | 71.6% | 27.7% | estimated ± 6.5 pp, low confidence |
| GPT-OSS 20B | 71.0% | 27.7% | estimated ± 6.5 pp, low confidence |
| Grok 4 | 72.6% | 27.7% | estimated ± 6.5 pp, low confidence |
| Kimi K2.5 | 77.2% | 27.9% | estimated ± 6.5 pp, low confidence |