integration bench
rev 01 · mean task score · public task setintegration bench
rev 01 · mean task score · public task set| # | Model | Task score | Score bar | Resolved | Status |
|---|---|---|---|---|---|
| 1 | 
Claude Fable 5 | 67.68 | 68% | ||
| 2 | 
GPT-5.6 Sol | 63.62 | 64% | ||
| 3 | 
DeepSeek V4 Pro (max) | 63.35 | 64% | ||
| 4 | 
Qwen 3.8 2.4T A95B | 61.68 | 62% | ||
| 5 | 
Grok 4.6 | 61.18 | 62% | ||
| 6 |
Claude Opus 5 | 59.32 | 60% | ||
| 7 |
Qwen 3.8 27B (xhigh) | 57.73 | 58% | ||
| 8 | 
Kimi K3 | 57.47 | 58% | ||
| 9 |
DeepSeek V4 Flash (max) | 55.68 | 56% | ||
| 10 |
Claude Opus 4.8 | 55.58 | 56% | ||
| 11 |
Claude Sonnet 5 | 53.85 | 54% | ||
| 12 |
GPT-5.6 Terra | 53.79 | 54% | ||
| 13 | 
Muse Spark 1.2 | 53.70 | 54% | ||
| 14 |
GPT-5.6 Luna | 51.78 | 52% | ||
| 15 | 
GLM 5.2 (xhigh) | 49.68 | 50% | ||
| 16 |
Claude Haiku 4.5 | 15.10 | 16% | ||
| 17 | 
Gemini 3.7 Flash | 35.85 | 36% | 28/50 graded |
This is the short version. Every number above comes from a recorded run. 170 of them are published, each with every tool call, every vendor request and the diff.
Replay them at integrationbench.com →