BlueFin - Interrogation: leaderboard

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; 36 held-out question-answering tasks over a workbook). Source: arxiv.org. Saturation forecast: Around March 2028. 5 models tracked.

Top models

#ModelScore
1GPT-5.5 (High)50
2Gemini 3.1 Pro (Preview) (High)44.4
3Claude Opus 4.7 (High)44.4
4Claude Sonnet 4.6 (High)40.3

Interactive version: theaggregate.ai/benchmark?slug=bluefin-interrogation · How It Works · Data refreshed daily, snapshot 2026-09-26.