BankerToolBench v1.1
Professional Work · 2026-10-09
Version 1.1 of BankerToolBench: the same 100 investment-banking workflows with renamed, stricter MCP tools, more indexed companies, WACC reference data exposed through the market-data tool, a clearer agent system prompt, and task specifications and rubrics revised by an extra layer of expert validation to be more objective. Agents run in OpenCode 1.17.9 under Harbor 0.15.0 with a Gemini 3 Flash verifier. A deliverable passes only if it meets every critical-weight (weight 10) rubric criterion, a stricter rule than v1.0's score threshold; the headline is Pass@1 averaged over replicate runs.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 32.6 |
| GPT-5.6 Sol | 29.0 |
| Opus 4.8 | 20.7 |
| GPT-5.5 | 15.0 |
| GPT-5.4 | 13.7 |
| Gemini 3.1 Pro Preview | 13.0 |
| GLM-5.2 | 12.4 |
| Opus 4.6 | 7.4 |
| GLM-5 | 4.7 |
| Qwen3.5 397B A17B | 2.7 |
| Gemini 2.5 Pro | 1.0 |