Hedge-Bench
Finance & Accounting · 2026-06-02
92 public-market research environments built from hedge fund analysts' recorded reasoning about one company, each paired with a later-resolved Yes/No question. Agents research frozen source files in a Harbor container without internet, write a cited report and forecast. Joint success needs a correct forecast and research coverage above 60% of the hidden analyst rubric, judged by a three-model jury; three trials per model.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 31.5 |
| Claude Opus 5 | 29.0 |
| Claude Fable 5.1 | 26.8 |
| Grok 4.6 | 26.4 |
| GLM-5.3 | 14.1 |
| GPT-5.6 Terra | 13.8 |
| Muse Spark 1.3 | 9.1 |
| Kimi K3 | 8.7 |
| Gemini 3.8 Flash | 4.7 |
| DeepSeek V4 Pro 0813 | 2.5 |