Vals AI Excel Modeling: leaderboard
Each task gives a business scenario and source files and asks an agent to build an Excel model, its tabs, formulas and formatting, across LBO, DCF, M&A, operating, comps and dataroom work; results use 51 held out tasks. Agents mainly use shell and recalculation tools. Template mode is graded by cell match against a gold workbook, scratch mode by rubric checks marked by a judge model, in percent, higher is better.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around January 2027. 76 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 | 76.67 | #9 |
| 2 | Claude Opus 5.5 | 75.94 | #3 |
| 3 | Claude Sonnet 5.5 | 75.71 | #11 |
| 4 | Gemini 4 Argon | 75.24 | #8 |
| 5 | Claude Fable 5 | 73.67 | #16 |
| 6 | Claude Opus 5 | 73.56 | #12 |
| 7 | GPT-5.6 Sol | 72.34 | #19 |
| 8 | Gemini 3.8 Flash | 72.2 | #21 |
| 9 | Muse Wonder Full | 71.92 | #10 |
| 10 | GPT-6 Astra | 71.7 | #6 |
| 11 | GPT-6 Sol | 71.53 | #14 |
| 12 | Gemini 3.7 Flash | 71.33 | #23 |
| 13 | GPT-6.1 Sol | 70.77 | #7 |
| 14 | Claude Opus 4.8 | 69.37 | #37 |
| 15 | GPT-6 Luna | 68.52 | #56 |
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-excel-modeling · How It Works · Data refreshed daily, snapshot 2026-10-20.