DI-Bench (Data Intelligence): leaderboard
DI-Bench (Zhang et al., 2026, Amazon) generates in-domain data-intelligence benchmarks for enterprise agents by linking relational databases to business-rule and metric documents; this board covers all 731 generated tasks over the Olist Brazilian e-commerce and PKDD 99 Czech banking datasets. An agent with SQL-execution and knowledge-base retrieval tools answers generated questions that need business rules and metric definitions applied to database computations, within 20 turns; exact match with task-specific tolerance, accuracy in percent, higher is better.
Metric: Accuracy (%; all 731 tasks: 449 Brazilian e-commerce (Olist) and 282 Czech retail-banking (PKDD 99) generated data-intelligence questions whose answers need retrieved business rules and metric definitions plus SQL over the database; agent with SQL-execution and knowledge-base tools, at most 20 turns, temperature 0, via the Amazon Bedrock Converse API; exact match with task-specific tolerance). Source: arxiv.org. Saturation forecast: Around May 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 57.9 | #100 |
| 2 | DeepSeek V3.2 | 50.9 | #403 |
| 3 | Claude Haiku 4.5 | 49 | #568 |
| 4 | Qwen 3 Next 80B A3B | 47.9 | #756 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=di-bench-data-intelligence · How It Works · Data refreshed daily, snapshot 2026-11-12.