Excel Modeling Benchmark v1.1
Professional Work · 2026-10-19
Excel Modeling Benchmark v1.1 reruns EMB's 51-task private test set with the GPT-6 Luna rubric judge and template-scoring fixes. Agents build investment-banking and private-equity Excel models (LBO, DCF, M&A, operating, comps, dataroom and miscellaneous) from a prompt and source files; template mode is graded by exact cell match against the gold workbook and scratch mode by a partial-credit rubric over numerical, formula and presentation checks. Kept separate from EMB v1 because the grading changed.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 76.7 |
| Gemini 4 Carbon | 76.5 |
| Claude Sonnet 5.5 | 75.9 |
| Claude Opus 5.5 | 75.8 |
| Muse Wonder Full | 75.0 |
| Gemini 4 Argon | 74.8 |
| GPT-6.1 Terra | 73.3 |
| Gemini 3.8 Flash | 72.9 |
| GPT-6 Astra | 72.7 |
| Muse Wonder | 72.5 |
| Qwen3.8 Max (0902) | 72.0 |
| Grok 4.7 | 71.5 |
| Kimi K3 | 71.3 |
| GPT-6.1 Sol | 71.1 |
| Kimi K3.1 Pro | 70.0 |
| GPT-6 Luna | 70.0 |
| Step 5 Preview | 69.6 |
| DeepSeek V4.1 Flash 1017 | 69.2 |
| GLM-5.3 | 69.1 |
| Mistral Large 4 Preview | 69.0 |
| Hy4 Preview | 68.9 |
| MiMo-V2.6-Flash | 68.9 |
| MiniMax M3.1 Flash | 67.8 |
| DeepSeek V4.1 Flash | 67.8 |
| MiMo-V2.6-Pro | 67.3 |
| GLM-5.3 Flash | 66.3 |
| Qwen3.8-Max | 66.0 |
| Claude Haiku 5.5 | 65.5 |
| DeepSeek V4 Pro 0813 | 65.1 |
| Beam 1.0 501B A23B | 64.7 |
| Muse Spark Open | 62.0 |
| MiniMax M3 | 59.8 |
| Mercury 3 | 57.9 |
| Inkling | 41.0 |
| Inkling-Small | 40.8 |