Calibration
Vals SWE-bench → NL2Repo
NL2Repo is estimated from Vals SWE-bench with a Hill curve fitted on 13 models measured on both: y = 0.3602 + (1.2000 − 0.3602)·x^6.00 / (1.12009^6.00 + x^6.00), R² = 0.76, cross-validated error 4.7 pp. It is used for 44 estimates.
| Estimated model | Vals SWE-bench | NL2Repo | Source |
|---|---|---|---|
| Claude Fable 5 | 95.0% | 58.8% | estimated ± 4.7 pp, high confidence |
| Claude Haiku 4.5 | 66.6% | 39.6% | estimated ± 4.7 pp, medium confidence |
| Claude Opus 4.7 | 82.0% | 47.2% | estimated ± 4.7 pp, high confidence |
| Claude Opus 4.8 | 88.6% | 52.5% | estimated ± 4.7 pp, high confidence |
| Claude Opus 5 | 97.0% | 60.9% | estimated ± 4.7 pp, medium confidence |
| Claude Sonnet 4.6 | 77.4% | 44.3% | estimated ± 4.7 pp, high confidence |
| Claude Sonnet 5 | 79.6% | 45.6% | estimated ± 4.7 pp, high confidence |
| Composer 2.5 | 79.6% | 45.6% | estimated ± 4.7 pp, high confidence |
| Gemini 2.5 Pro | 54.4% | 37.1% | estimated ± 4.7 pp, medium confidence |
| Gemini 3.1 Flash-Lite | 62.8% | 38.5% | estimated ± 4.7 pp, medium confidence |
| Gemini 3.1 Pro | 78.8% | 45.1% | estimated ± 4.7 pp, high confidence |
| Gemini 3.5 Flash | 78.8% | 45.1% | estimated ± 4.7 pp, high confidence |
| Gemini 3.5 Flash-Lite | 75.0% | 43.0% | estimated ± 4.7 pp, high confidence |
| Gemini 3.6 Flash | 79.6% | 45.6% | estimated ± 4.7 pp, high confidence |
| Gemini 3.7 Flash | 80.8% | 46.4% | estimated ± 4.7 pp, high confidence |
| Gemini 3.8 Flash | 80.0% | 45.9% | estimated ± 4.7 pp, high confidence |
| Gemini 3 Flash | 75.0% | 43.0% | estimated ± 4.7 pp, high confidence |
| GLM-4.7 | 69.4% | 40.5% | estimated ± 4.7 pp, high confidence |
| GPT-5.2-Codex | 72.4% | 41.7% | estimated ± 4.7 pp, high confidence |
| GPT-5.3 Codex | 78.0% | 44.6% | estimated ± 4.7 pp, high confidence |
| GPT-5.4 mini | 73.0% | 42.0% | estimated ± 4.7 pp, high confidence |
| GPT-5.4 nano | 69.8% | 40.7% | estimated ± 4.7 pp, high confidence |
| GPT-5.5 | 82.6% | 47.7% | estimated ± 4.7 pp, high confidence |
| GPT-5.6 Luna | 93.0% | 56.7% | estimated ± 4.7 pp, high confidence |
| GPT-5.6 Sol | 96.2% | 60.1% | estimated ± 4.7 pp, high confidence |
| GPT-5.6 Terra | 95.4% | 59.2% | estimated ± 4.7 pp, high confidence |
| Grok 4.20 | 72.2% | 41.6% | estimated ± 4.7 pp, high confidence |
| Grok 4.3 | 71.4% | 41.3% | estimated ± 4.7 pp, high confidence |
| Grok 4.5 | 86.6% | 50.8% | estimated ± 4.7 pp, high confidence |
| Grok 4.6 | 95.6% | 59.4% | estimated ± 4.7 pp, high confidence |
| Inkling | 77.6% | 44.4% | estimated ± 4.7 pp, high confidence |
| Inkling-Small | 82.2% | 47.4% | estimated ± 4.7 pp, high confidence |
| Kimi K2.7 Code | 78.2% | 44.7% | estimated ± 4.7 pp, high confidence |
| Kimi K3 | 93.4% | 57.1% | estimated ± 4.7 pp, high confidence |
| Laguna M.1 | 57.6% | 37.5% | estimated ± 4.7 pp, medium confidence |
| Laguna XS.2 | 55.2% | 37.2% | estimated ± 4.7 pp, medium confidence |
| Ling 3.0 Flash | 65.2% | 39.2% | estimated ± 4.7 pp, medium confidence |
| MiMo-V2.5 | 71.0% | 41.1% | estimated ± 4.7 pp, high confidence |
| MiMo-V2.5-Pro | 74.0% | 42.5% | estimated ± 4.7 pp, high confidence |
| Mistral Medium 3.5 128B | 66.4% | 39.5% | estimated ± 4.7 pp, medium confidence |
| Muse Spark | 74.4% | 42.7% | estimated ± 4.7 pp, high confidence |
| Muse Spark 1.1 | 82.0% | 47.2% | estimated ± 4.7 pp, high confidence |
| Muse Spark 1.2 | 86.6% | 50.8% | estimated ± 4.7 pp, high confidence |
| Qwen3.5 Flash | 64.4% | 38.9% | estimated ± 4.7 pp, medium confidence |