BlueFin - Public Set: leaderboard
The 11 publicly released tasks of BlueFin (Kundurthy et al., 2026): 7 manipulation tasks, 1 synthesis task and 3 interrogation questions over financial-model workbooks, solved by an agent through a 20-tool spreadsheet harness. Rubric-weighted score in percent, averaged over the 11 tasks (higher is better); rubric criteria are checked by an agentic GPT-5.4 judge.
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; micro-average over the 11 released tasks: 7 manipulation, 1 synthesis, 3 interrogation). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (High) | 58.4 | #75 (Claude Opus 4.7) |
| 2 | GPT-5.5 (High) | 54.1 | #48 (GPT-5.5) |
| 3 | Claude Sonnet 4.6 (High) | 48.7 | #158 (Claude Sonnet 4.6) |
| 4 | Gemini 3.1 Pro (Preview) (High) | 46.8 | #91 (Gemini 3.1 Pro (Preview)) |
| 5 | Kimi K2.6 | 42.3 | #173 |
| 6 | Grok 4.20 (High) | 23.3 | #296 (Grok 4.20) |
| 7 | MiniMax-M2.5 | 20.9 | #590 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bluefin-public-set · How It Works · Data refreshed daily, snapshot 2026-11-12.