Business Bench - Desk 2026-09-28 - Spreadsheet: leaderboard
Business Bench's September 28, 2026 desk campaign Spreadsheet compares three configured agent systems on 187 public business-deliverable tasks repeated three times. Executable conservative-v7 checks grade the produced files; this panel reports Frozen-Grader Attempt Acceptance (%). Model, harness build, provider and sampling differ and remain explicit. Tasks were used during Proto development and the frozen scorer was selected after outputs existed; published results describe this dated task cohort.
Metric: Frozen-Grader Attempt Acceptance (%). Source: github.com. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Sol (Business Bench Run: desk 2026-09-28; proto-sol6-sub; high; 0d09882b59a1) | 97.56 | |
| 2 | DeepSeek V4.1 Flash (Business Bench Run: desk 2026-09-28; proto-deepseek-direct; high; 0d09882b59a1) | 94.31 | |
| 3 | GPT-6 Sol (Business Bench Run: desk 2026-09-28; codex-sol6; high; 0d09882b59a1) | 93.5 |
Interactive version: theaggregate.ai/benchmark?slug=business-bench-desk-2026-09-28-spreadsheet · How It Works · Data refreshed daily, snapshot 2026-11-05.