Calibration
Vibe Code Bench → PostTrainBench v1.1
PostTrainBench v1.1 is estimated from Vibe Code Bench with a Hill curve fitted on 5 models measured on both: y = 0.1985 + (1.2000 − 0.1985)·x^6.00 / (1.10324^6.00 + x^6.00), R² = 0.89, cross-validated error 10.5 pp. It is used for 16 estimates.
| Estimated model | Vibe Code Bench | PostTrainBench v1.1 | Source |
|---|---|---|---|
| Claude Haiku 4.5 Thinking | 11.4% | 19.9% | estimated ± 10.5 pp, low confidence |
| Claude Opus 4.5 Thinking | 20.6% | 19.9% | estimated ± 10.5 pp, low confidence |
| Claude Opus 4.6 (Adaptive) | 53.5% | 21.1% | estimated ± 10.5 pp, low confidence |
| Claude Sonnet 4.5 Thinking | 22.6% | 19.9% | estimated ± 10.5 pp, low confidence |
| DeepSeek V3.2 (Thinking) | 5.1% | 19.9% | estimated ± 10.5 pp, low confidence |
| Gemini 3 Pro | 14.3% | 19.9% | estimated ± 10.5 pp, low confidence |
| GLM-4.6 | 3.1% | 19.9% | estimated ± 10.5 pp, low confidence |
| GLM-5 (Reasoning) | 23.4% | 19.9% | estimated ± 10.5 pp, low confidence |
| GPT-5.1-Codex | 13.1% | 19.9% | estimated ± 10.5 pp, low confidence |
| GPT-5.1-Codex-Max | 22.2% | 19.9% | estimated ± 10.5 pp, low confidence |
| GPT-5 mini | 14.2% | 19.9% | estimated ± 10.5 pp, low confidence |
| Grok 4.1 Fast (Reasoning) | 1.2% | 19.9% | estimated ± 10.5 pp, low confidence |
| Grok 4 Fast (Reasoning) | 0.0% | 19.9% | estimated ± 10.5 pp, low confidence |
| MiniMax M2.5 | 14.9% | 19.9% | estimated ± 10.5 pp, low confidence |
| Qwen3.5 Plus | 15.7% | 19.9% | estimated ± 10.5 pp, low confidence |
| Qwen3 Max | 3.5% | 19.9% | estimated ± 10.5 pp, low confidence |