BankerToolBench
Professional Work · 2026-04-13
Version 1.0 of BankerToolBench (BTB), Handshake AI's benchmark of 100 end-to-end junior investment-banking workflows built with 502 bankers. Agents in the OpenCode harness use SEC EDGAR, market-data and brand-asset MCP tools to produce Excel models, PowerPoint decks and Word or PDF memos, which the Gandalf agentic verifier grades against banker-written rubrics averaging about 150 binary criteria weighted 1, 3, 5 or 10. The headline Score is the weighted percentage of criteria passed, averaged over tasks and three runs; the paper recommends it for fine-grained model comparison.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.4 | 58.1 |
| GPT-5.2 | 56.1 |
| Gemini 3.1 Pro Preview | 53.6 |
| Opus 4.6 | 53.2 |
| Opus 4.5 | 52.3 |
| GLM-5 | 46.8 |
| Qwen3.5 397B A17B | 42.6 |
| Grok 4 | 31.4 |
| Gemini 2.5 Pro | 29.4 |