Toolathlon: leaderboard
The Tool Decathlon sets an agent 108 tasks on real software, spanning 32 applications and 604 tools from Google Calendar to Kubernetes, at about twenty tool-calling turns per task. Dedicated scripts grade each task by checking the final environment state; figures are from benchmarklist.com, many self-reported and given as Pass@1 where the row names a metric, and higher is better. Source: toolathlon.xyz/introduction.
Metric: Score (mixed sources). Source: toolathlon.xyz. Source: benchmarklist.com. Saturation forecast: Around March 2028. 54 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.3 Flash | 78.4 | #65 |
| 2 | Claude Fable 5 | 77.9 | #16 |
| 3 | Claude Opus 5 | 77.6 | #12 |
| 4 | Kimi K3 | 76.5 | #30 |
| 5 | Kimi K3 (Max) | 76.5 | #30 (Kimi K3) |
| 6 | Claude Opus 4.8 | 76.2 | #37 |
| 7 | DeepSeek-V4-Flash-Vision-Exp | 75.9 | #76 |
| 8 | Muse Spark 1.1 | 75.6 | #52 |
| 9 | GPT-5.6 Sol | 74.9 | #19 |
| 10 | GPT-5.6 Terra | 74.9 | #41 |
| 11 | DeepSeek V4 Pro (0813) | 74.1 | #53 |
| 12 | GPT-5.5 (xHigh) | 73.5 | #31 (GPT-5.5) |
| 13 | Qwen3.8-Flash-Next | 73.5 | #72 |
| 14 | Qwen 3.8 Max (0902) | 73.3 | #29 |
| 15 | GLM-5.3 | 73 | #39 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=toolathlon · How It Works · Data refreshed daily, snapshot 2026-10-20.