Ramp Accounting Bench
Finance & Accounting
Ramp Labs' 137 long-horizon accounting tasks across 22 simulated companies (QuickBooks ledgers, bank statements, invoices, contracts, reconciliation workbooks), solved with 92 MCP domain tools in an Archipelago-style ReAct loop. Mean@3 averages the fraction of human-written rubric criteria a GPT-5.6 Luna judge marks met, over three attempts per task. A separate 137-task cohort from APEX-Accounting's 160 tasks.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Argon | 61.1 |
| Claude Opus 5.5 | 50.5 |
| Claude Fable 5.1 | 49.7 |
| GPT-6 Astra | 48.7 |
| GPT-6.1 Sol | 46.4 |
| Gemini 3.8 Flash | 45.7 |
| Grok 4.7 | 44.1 |
| GLM-5.3 | 43.1 |
| Claude Sonnet 5.5 | 42.5 |
| Gemini 3.7 Flash | 42.3 |
| Claude Haiku 5.5 | 42.3 |
| Qwen3.8-Max | 41.2 |
| DeepSeek V4.1 Flash | 40.8 |
| Kimi K3 | 39.2 |
| GLM-5.3 Flash | 35.7 |
| GPT-6 Luna | 34.0 |