CRM-Agent-Bench
Agents · 2026-10-22
555 stateful, multi-turn customer relationship management tasks across nine business domains and 36 capability tags from Salesforce. A simulated user (gpt-5-mini) talks to the agent, which acts through typed CRM tools; each task carries an executable contract over actions, arguments, order, response, target state and protected state, and strict success requires every declared check. One trial per task; the score is the share of the 555 tasks passed, in percent.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.8 Flash | 83.6 |
| Claude Opus 5 | 82.7 |
| Sonnet 4.6 | 82.0 |
| GPT-5.6 Terra | 81.6 |
| Sonnet 5 | 81.3 |
| GPT-5.5 | 81.3 |
| GPT-5.6 Sol | 81.3 |
| GPT-5.6 Luna | 80.7 |
| Gemini 3.1 Pro Preview | 79.1 |
| GPT-5.4 Nano | 79.1 |
| Muse Glimmer 30B | 78.7 |
| Haiku 4.5 | 76.0 |
| Nemotron 3 Super 120B A12B | 65.6 |
| GPT-4.1 Nano | 64.3 |
| Nemotron 3.5 Lightning 30B A3B (precision unspecified) | 64.3 |