Performance and cost by setup
PenguinHarness · DeepSeek V4 Flash
Claude Code · Claude Opus 4.8
Codex · GPT-5.5
Data Analysis Bench
A benchmark for end-to-end data analysis.
15 tasks covering documents, spreadsheets, databases, multi-source analysis, and delivery.
- 15
- Cases
- Complex data-analysis tasks spanning documents, spreadsheets, databases, and multi-source workflows.
- 8
- Settings
- Multiple PenguinHarness setups, plus Claude Code and Codex.
- 11/15
- Best result
- PenguinHarness v0.1.5 manual tuning.
Results
15-case results
One run per setting. Accuracy is the PASS rate.
| Loading published results… |
Benchmark coverage
Core challenges in end-to-end data analysis.
Long-document retrieval, complex tables, cross-source verification, analytical reasoning, and delivery.
Documents & evidence
Retrieve and bind evidence across long documents, scanned pages, and multi-document collections.
Spreadsheets & table transformation
Work with hierarchical tables, formulas, workbook models, and schema normalization.
Databases & multi-source workflows
Query relational data, navigate workspace permissions, and reconcile information across sources.
Analytical reasoning & visualization
Compute derived metrics and produce or validate analytical visuals.