Calibration
Vibe Code Bench → AA-SciCode
AA-SciCode is estimated from Vibe Code Bench with a linear curve fitted on 12 models measured on both: y = 0.1451·x + 0.4574, R² = 0.54, cross-validated error 3.8 pp. It is used for 19 estimates.
| Estimated model | Vibe Code Bench | AA-SciCode | Source |
|---|---|---|---|
| Claude Haiku 4.5 Thinking | 11.4% | 47.4% | estimated ± 3.8 pp, high confidence |
| Claude Opus 4.5 Thinking | 20.6% | 48.7% | estimated ± 3.8 pp, high confidence |
| Claude Opus 4.6 (Adaptive) | 53.5% | 53.5% | estimated ± 3.8 pp, high confidence |
| Claude Sonnet 4.5 Thinking | 22.6% | 49.0% | estimated ± 3.8 pp, high confidence |
| DeepSeek V3.2 (Thinking) | 5.1% | 46.5% | estimated ± 3.8 pp, high confidence |
| Gemini 3.1 Flash-Lite | 0.0% | 45.7% | estimated ± 3.8 pp, medium confidence |
| Gemini 3 Flash | 20.2% | 48.7% | estimated ± 3.8 pp, high confidence |
| Gemini 3 Pro | 14.3% | 47.8% | estimated ± 3.8 pp, high confidence |
| GLM-4.6 | 3.1% | 46.2% | estimated ± 3.8 pp, high confidence |
| GLM-5 (Reasoning) | 23.4% | 49.1% | estimated ± 3.8 pp, high confidence |
| GPT-5.1-Codex | 13.1% | 47.6% | estimated ± 3.8 pp, high confidence |
| GPT-5.1-Codex-Max | 22.2% | 49.0% | estimated ± 3.8 pp, high confidence |
| GPT-5.2-Codex | 37.9% | 51.2% | estimated ± 3.8 pp, high confidence |
| GPT-5 mini | 14.2% | 47.8% | estimated ± 3.8 pp, high confidence |
| Grok 4.1 Fast (Reasoning) | 1.2% | 45.9% | estimated ± 3.8 pp, high confidence |
| Grok 4 Fast (Reasoning) | 0.0% | 45.7% | estimated ± 3.8 pp, medium confidence |
| MiniMax M2.5 | 14.9% | 47.9% | estimated ± 3.8 pp, high confidence |
| Qwen3.5 Plus | 15.7% | 48.0% | estimated ± 3.8 pp, high confidence |
| Qwen3 Max | 3.5% | 46.2% | estimated ± 3.8 pp, high confidence |