Leaderboard
BlueFin
LLM agents on professional-finance spreadsheet work: building new workbooks (synthesis), editing existing ones (manipulation), and answering questions about them (interrogation/comprehension). It puts particular weight on 'dynamic correctness', meaning formulas that still work when inputs change.
BlueFin arXiv v1 (2026-05-29)
criteria-weighted score % (overall), higher is better. 5 models.
Updated May 2026
- GPT 5.549.6
- Opus 4.749.2
- gemini-3.1-pro45.9
- Sonnet 4.642.9
- Grok 4.2031.3
BlueFin arXiv v1 (2026-05-29)
criteria-weighted score % (overall), higher is better. 5 models.
Updated May 2026
- 1GPT 5.549.6
- 2Opus 4.749.2
- 3gemini-3.1-pro45.9
- 4Sonnet 4.642.9
- 5Grok 4.2031.3
| 1 | GPT 5.5 | OpenAI | 49.6 | held-out set, 120 tasks | Paper Table 2 Reported by the authors, May 2026 |
|---|---|---|---|---|---|
| 2 | Opus 4.7 | Anthropic | 49.2 | held-out set, 120 tasks | Paper Table 2 Reported by the authors, May 2026 |
| 3 | gemini-3.1-pro | 45.9 | held-out set, 120 tasks | Paper Table 2 Reported by the authors, May 2026 | |
| 4 | Sonnet 4.6 | Anthropic | 42.9 | held-out set, 120 tasks | Paper Table 2 Reported by the authors, May 2026 |
| 5 | Grok 4.20 | xAI | 31.3 | held-out set, 120 tasks | Paper Table 2 Reported by the authors, May 2026 |