Stepfun; FinStep, 2025
FinResearchBench
Measures how well deep-research agents write financial research outputs (stock, event, sector, morning brief, post-market recap and similar). An Agent-as-a-Judge first extracts a logic tree of arguments and evidence, then scores the output against it.
About this benchmark
- Size
- 70 questions (10 for each of 7 task types)
- Public split
- none (no data or code linked)
- Grading
- Ten dimensions, each scored 0-100 and weighted equally into a Final Score. Six are scored by an LLM with a rubric (Claude 3.7 Sonnet is the core judge) and four are computed by rules.
- Licence
- Not stated
- Submission
- none
- Products allowed
- Yes
- Latest version
- arXiv v3 (2025-10-20)
- What it does not cover
- Does not check numeric facts against sources or forecast accuracy. It scores structure and argument quality.
Results
8 modelsEach score shows who reported it: the benchmark's authors (quoted from their paper), the benchmark's official board, or Moneda (run with the benchmark's own grader and rules). We do not copy other aggregators.
FinResearchBench arXiv v3 (2025-10-20)
Final Score (0-100), higher is better. 8 models.
Updated 20 Oct 2025
- Gemini76.6
- OpenAI74.2
- Cashcat Pro70.7
- Minimax*67.7
- Reportify67.2
- Perplexity67.0
- Genspark*64.3
- AutoGLM60.0
FinResearchBench arXiv v3 (2025-10-20)
Final Score (0-100), higher is better. 8 models.
Updated 20 Oct 2025
- 1Gemini76.6
- 2OpenAI74.2
- 3Cashcat Pro70.7
- 4Minimax*67.7
- 5Reportify67.2
- 6Perplexity67.0
- 7Genspark*64.3
- 8AutoGLM60.0
| 1 | Gemini | 76.6 | all 70 questions | Paper Table 1 Reported by the authors, 20 Oct 2025 | |
|---|---|---|---|---|---|
| 2 | OpenAI | OpenAI | 74.2 | all 70 questions | Paper Table 1 Reported by the authors, 20 Oct 2025 |
| 3 | Cashcat Pro | Not stated | 70.7 | all 70 questions | Paper Table 1 Reported by the authors, 20 Oct 2025 |
| 4 | Minimax* | MiniMax | 67.7 | all 70 questions | Paper Table 1 Reported by the authors, 20 Oct 2025 |
| 5 | Reportify | Not stated | 67.2 | all 70 questions | Paper Table 1 Reported by the authors, 20 Oct 2025 |
| 6 | Perplexity | Perplexity | 67.0 | all 70 questions | Paper Table 1 Reported by the authors, 20 Oct 2025 |
| 7 | Genspark* | Not stated | 64.3 | all 70 questions | Paper Table 1 Reported by the authors, 20 Oct 2025 |
| 8 | AutoGLM | Zhipu AI | 60.0 | all 70 questions | Paper Table 1 Reported by the authors, 20 Oct 2025 |