Realm Financial Reasoning
Finance & Accounting
micro1's Realm finance-reasoning benchmark: 103 expert-authored tasks built on practitioners' work product (IFRS reconciliations, hedge-fund backtests, venture term sheets, treasury forecasts and more) with 144 source attachments. An agent works each task in a fresh OpenCode sandbox with a shell, file tools, web search and web fetch, and a GPT-5.5 mini judge grades the answer against a weighted rubric of positive and negative criteria. A run's reward is earned weight over total positive weight, clamped to 0-1, and the score is the mean reward.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.8 Flash | 51.5 |
| Claude Fable 5.1 | 47.8 |
| Claude Opus 5 | 47.4 |
| Kimi K3 | 45.7 |
| Muse Spark 1.3 | 45.5 |
| Muse Spark 1.1 | 45.0 |
| Grok 4.6 | 44.6 |
| Grok 4.7 | 44.5 |
| GPT-5.6 Sol | 44.4 |
| GPT-5.5 | 44.3 |
| Sonnet 5 | 43.2 |
| Opus 4.8 | 43.1 |
| Grok 4.5 | 43.1 |
| Gemini 3.1 Pro Preview | 39.6 |
| Inkling | 34.6 |