FinLifeBench - Financial State - Checkpoint State Accuracy: leaderboard
FinLifeBench (arXiv 2609.01198v2, Table 3): 11 models reconstruct customer records from up to 300 cumulative Korean banking sessions per persona (20 personas, 400 checkpoints per task), full context, low reasoning setting where available. Cell-level accuracy of the predicted value and validity status for all 34 state paths at each checkpoint; percent, higher is better.
Metric: Cell-level state accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 (Low) | 80.1 | #39 (Claude Opus 4.8) |
| 2 | GPT-5.6 Luna (Low) | 77.1 | #79 (GPT-5.6 Luna) |
| 3 | Gemini 3.1 Pro (Preview) (Low) | 76.7 | #64 (Gemini 3.1 Pro (Preview)) |
| 4 | Gemini 3.5 Flash (Low) | 75.4 | #66 (Gemini 3.5 Flash) |
| 5 | GPT-5.6 Sol (Low) | 74.9 | #22 (GPT-5.6 Sol) |
| 6 | GPT-5.6 Terra (Low) | 74.7 | #40 (GPT-5.6 Terra) |
| 7 | Claude Sonnet 4.6 (Low) | 71.1 | #94 (Claude Sonnet 4.6) |
| 8 | GPT-OSS-120B (Low) | 67.4 | #346 (GPT-OSS-120B) |
| 9 | Llama 4 Maverick | 67.2 | #486 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=finlifebench-financial-state-checkpoint-state-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.