GDPevo: leaderboard
Agents perform stateful business tasks across CRM, ERP, finance, healthcare, legal work, data analysis and engineering operations: update records, inspect evidence, perform calculations and produce checked deliverables. Solves held-out tasks directly without learning from the training tasks. Current released 24-group evaluation uses 120 held-out tasks from 240 total train/test tasks, with three repetitions. Native Codex or Claude Code harness and source xhigh/max reasoning are preserved. Current README/frontend leaderboard is used; archived 12-group configurations are not substituted.
Metric: Mean accuracy (%), base, 24 task groups, three repetitions. Source: github.com. Saturation forecast: Around June 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 (xHigh) | 50.63 | #39 (Claude Opus 4.8) |
| 2 | GPT-5.5 (xHigh) | 49.37 | #33 (GPT-5.5) |
| 3 | GLM-5.2 (Max) | 46.12 | #80 (GLM-5.2) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=gdpevo · How It Works · Data refreshed daily, snapshot 2026-11-05.