Analytics agents
AIDABench
Spreadsheet, report, chart, and file-generation analytics tasks with separate evaluators.
As of 13 August 2026. Held out. Yang et al., arXiv 2603.15636. No explicit data license. The engine catalog blocks it.
AIDABench tests analytics agents on spreadsheets, reports, charts, and generated files. Pass-at-1, visualization checks, and file checks are different gates. The paper's best reported pass-at-1 is 59.4%.
Until a data license is obtained, the pack cannot enter a commercial artifact. The 59.4% cell is the paper headline, not a 2026 Fable 5 or GPT-5.6 replay.
- Scale
- The paper reports a best pass-at-1 of 59.4%
- Provenance
- Yang et al., arXiv 2603.15636. No explicit data license at the official repository. Blocked in the engine catalog.
- Contamination
- Medium. Public since 2026.
- Cannot establish
- Any commercial client or pitch artifact until a data license is obtained.
Metrics, kept separate
- Pass-at-1
- Answer checks
- Visualization checks
- File-output checks
Paper table · cited paper · not an Antidote run · not a product score
Yang et al., arXiv 2603.15636. Best reported pass-at-1. Not a Fable 5 or GPT-5.6 cell. Pack remains held out.
| System | Pass-at-1 |
|---|---|
| Paper best reported pass-at-1 | 59.4% |
What the figures bound
The pack is listed because spreadsheet and file-generation work is economically material. Saturation status is not permission to run it.
Do not treat 59.4% as a current frontier VLM score.
What this page does not conclude
- A commercial or pitch use of AIDABench
- GPT-5.6 or Fable 5 analytics-agent performance
- Spreadsheet-product substitution
- An overall analytics winner
Cited as calibration on these memos — not a score of those products.
Not attached to a software, advisory, or automation memo yet.