WorkBench 2026: leaderboard
Metric: Successful task completion (%) on the corrected 2026 release of WorkBench: 690 templated workplace tasks over a sandbox of five databases (calendar, email, website analytics, CRM, project board) and 26 tools, graded by comparing the final sandbox state with the ground truth; one native tool-calling agent loop for every model, up to 20 steps, temperature 0 where allowed; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 24 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5 | 97.7 |
| 2 | Claude Opus 4.8 | 96.2 |
| 3 | Gemini 3.1 Pro (Preview) | 95.5 |
| 4 | GPT-5.5 | 94.9 |
| 5 | Gemini 3.5 Flash | 92 |
| 6 | Claude Sonnet 4.6 | 88.3 |
| 7 | Kimi K2.6 | 88 |
| 8 | GPT-5 | 84.3 |
| 9 | DeepSeek V4 Pro | 84.2 |
| 10 | GPT-5.4 | 78.6 |
| 11 | O3 | 77.4 |
| 12 | GLM-4.6 | 77.4 |
| 13 | Claude Haiku 4.5 | 74.8 |
| 14 | GPT-4.1 | 73.5 |
| 15 | GPT-5.2 | 70.4 |
Interactive version: theaggregate.ai/benchmark?slug=workbench-2026 · How It Works · Data refreshed daily, snapshot 2026-09-29.