FRTR-Bench: leaderboard
Spreadsheet question answering over 30 large synthetic enterprise Excel workbooks (FRTR-Bench, Gulati et al., 2026): 157 questions need lookups, aggregation, cross-sheet formulas or reading embedded charts and receipts, across workbooks of up to 210,000 rows. Every model answers through the same retrieval pipeline (the paper's FRTR: row, column, window and image chunks, hybrid BM25 and dense retrieval with reciprocal rank fusion), and an answer counts if it is numerically consistent with the reference or an equivalent formula. Accuracy in percent; higher is better.
Metric: Answer accuracy (%; 157 questions over 30 workbooks, answered through the FRTR hybrid-retrieval pipeline). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 74 | #244 |
| 2 | GPT-5 | 73 | #163 |
| 3 | Gemini 2.5 Pro | 67 | #262 |
| 4 | Llama 4 Maverick Instruct | 56 | #1125 |
| 5 | GPT-OSS-120B | 51 | #699 |
| 6 | GPT-4o | 49 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=frtr-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.