BlueFin: leaderboard
BlueFin (Kundurthy et al., 2026) evaluates LLM agents on professional-finance spreadsheet work through a 20-tool spreadsheet harness: building models from a brief, modifying existing models and answering questions about the calculations in a workbook. Each task scores the weighted share of expert-validated rubric points met, checked by an agentic GPT-5.4 judge that can change inputs and recalculate, or the correctness of the answers; the board averages the 120 held-out tasks in percent, with non-completions scored zero (higher is better).
Metric: Criteria-weighted rubric score (%; points met over maximum points of expert rubrics checked by an agentic GPT-5.4 judge for synthesis and manipulation workbooks, answer correctness for interrogation, non-completions count as zero; 20-tool spreadsheet harness, high reasoning effort; 120 held-out tasks: 75 manipulation, 9 synthesis, 36 interrogation). Source: arxiv.org. Saturation forecast: Around August 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (High) | 49.6 | #48 (GPT-5.5) |
| 2 | Claude Opus 4.7 (High) | 49.2 | #75 (Claude Opus 4.7) |
| 3 | Gemini 3.1 Pro (Preview) (High) | 45.9 | #91 (Gemini 3.1 Pro (Preview)) |
| 4 | Claude Sonnet 4.6 (High) | 42.9 | #158 (Claude Sonnet 4.6) |
| 5 | Grok 4.20 (High) | 31.3 | #296 (Grok 4.20) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bluefin · How It Works · Data refreshed daily, snapshot 2026-11-12.