Calibration
GDPval-AA → OSWorld-Verified
OSWorld-Verified is estimated from GDPval-AA with a linear curve fitted on 21 models measured on both: y = 0.5293·x + 0.5541, R² = 0.49, cross-validated error 8.4 pp. It is used for 58 estimates.
| Estimated model | GDPval-AA | OSWorld-Verified | Source |
|---|---|---|---|
| Claude Opus 5.5 | 68.3% | 91.6% | estimated ± 8.4 pp, low confidence |
| Muse Wonder Full | 67.9% | 91.3% | estimated ± 8.4 pp, low confidence |
| Claude Sonnet 5.5 | 66.9% | 90.8% | estimated ± 8.4 pp, low confidence |
| Muse Wonder | 63.5% | 89.0% | estimated ± 8.4 pp, low confidence |
| Claude Fable 5.1 | 62.7% | 88.6% | estimated ± 8.4 pp, low confidence |
| Grok 4.7 | 60.7% | 87.5% | estimated ± 8.4 pp, low confidence |
| Qwen3.8 Max Preview | 58.3% | 86.3% | estimated ± 8.4 pp, low confidence |
| Gemini 4 Argon | 56.2% | 85.2% | estimated ± 8.4 pp, low confidence |
| Claude Haiku 5.5 | 56.0% | 85.1% | estimated ± 8.4 pp, low confidence |
| GPT-6.1 Sol | 54.8% | 84.4% | estimated ± 8.4 pp, low confidence |
| GPT-6 Sol | 50.4% | 82.1% | estimated ± 8.4 pp, low confidence |
| Muse Spark 1.2 | 48.6% | 81.1% | estimated ± 8.4 pp, low confidence |
| GPT-6 Luna | 46.6% | 80.1% | estimated ± 8.4 pp, low confidence |
| Muse Spark Open | 45.1% | 79.3% | estimated ± 8.4 pp, low confidence |
| Grok 4.5 | 44.2% | 78.8% | estimated ± 8.4 pp, low confidence |
| Apodex 1.1 Mini | 34.4% | 73.6% | estimated ± 8.4 pp, low confidence |
| Quasar 438B | 34.1% | 73.5% | estimated ± 8.4 pp, low confidence |
| Ling 3.0 Flash VL | 33.0% | 72.9% | estimated ± 8.4 pp, low confidence |
| Hy3 | 28.2% | 70.3% | estimated ± 8.4 pp, low confidence |
| Kimi K2.7 Code | 26.9% | 69.6% | estimated ± 8.4 pp, low confidence |
| A.X K2 | 23.0% | 67.6% | estimated ± 8.4 pp, low confidence |
| GPT-5 | 20.9% | 66.5% | estimated ± 8.4 pp, low confidence |
| MiniCPM5-2B | 10.4% | 60.9% | estimated ± 8.4 pp, low confidence |
| MiMo-V2-Flash | 6.4% | 58.8% | estimated ± 8.4 pp, low confidence |
| Granite 4.2 30B | 3.7% | 57.4% | estimated ± 8.4 pp, low confidence |
| Gemma 4 26B A4B | 3.2% | 57.1% | estimated ± 8.4 pp, low confidence |
| Ling 3.0 Tiny | 3.0% | 57.0% | estimated ± 8.4 pp, low confidence |
| Command A+ | 0.7% | 55.8% | estimated ± 8.4 pp, low confidence |
| Celeris-1 | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Claude 3.5 Haiku | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| DeepSeek V3 | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| DeepSeek V3 0324 | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Gemma 3 27B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Gemma 4 12B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Gemma 4 E2B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Gemma 4 E4B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| GPT-4.1 mini | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| GPT-4.1 nano | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| GPT-4o | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| GPT-4o mini | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| GPT-OSS 20B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Granite 4.2 3B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Granite 4.2 8B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| K-Exaone | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Ling 2.6 Flash | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Llama 3.3 70B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Llama 4 Maverick | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Llama 4 Scout | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Mercury 2.5 | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Mistral Large 3 | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Mistral Small 3.1 | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Mistral Small 4 | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Nemotron 3 Nano 30B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Nemotron 3 Nano Omni 30B A3B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| North Mini Code | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Solar Pro 3 | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Trinity-Large-Preview | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
| Ultravox v0.6 Llama 3.3 70B | 0.0% | 55.4% | estimated ± 8.4 pp, low confidence |
Fitted on
| Model | GDPval-AA | OSWorld-Verified |
|---|---|---|
| Claude Fable 5 | 55.5% | 85.0% |
| Claude Opus 4.7 | 42.6% | 78.0% |
| Claude Opus 4.8 | 47.7% | 83.4% |
| Claude Sonnet 5 | 48.2% | 81.2% |
| Gemini 3.5 Flash | 42.2% | 78.4% |
| Gemini 3.5 Flash-Lite | 24.2% | 74.0% |
| Gemini 3.6 Flash | 39.2% | 83.0% |
| GPT-5.4 | 37.3% | 75.0% |
| GPT-5.4 mini | 25.6% | 72.1% |
| GPT-5.4 nano | 22.4% | 39.0% |
| GPT-5.5 | 42.5% | 78.7% |
| Kimi K2.6 | 26.9% | 73.1% |
| Kimi K2.5 | 17.0% | 63.3% |
| MiMo-V2.6-Flash | 55.4% | 80.8% |
| MiMo-V2.6-Pro | 59.2% | 82.0% |
| MiniMax M3 | 37.1% | 70.1% |
| Muse Glimmer 30B | 14.2% | 65.9% |
| Muse Spark 1.1 | 35.6% | 80.8% |
| Qwen3.5-122B-A10B | 15.7% | 58.0% |
| Qwen3.7 Plus | 13.3% | 73.3% |
| Qwen3.8-27B | 45.9% | 84.3% |