# Deep FinResearch Bench leaderboard

Scores equity research reports from deep-research agents against professional analyst reports on three axes: qualitative rigour, forecast and valuation accuracy (SMAPE, target-price hit rate, recommendation direction), and how credible and verifiable each claim is.

## Deep FinResearch Bench: overall factual claim rate % (higher is better, 4 models)

| Rank | Model | Maker | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | OpenAI | OpenAI | 86 | overall factual claim rate % | OpenAI DR (GPT-o3), Firm A-structured reports | Reported by the authors | Paper Table 7 | https://arxiv.org/html/2604.21006v1 | 22 Apr 2026 |
| 2 | Perplexity | Perplexity | 75.6 | overall factual claim rate % | Perplexity DR (Sonar Reasoning), Firm A-structured reports | Reported by the authors | Paper Table 7 | https://arxiv.org/html/2604.21006v1 | 22 Apr 2026 |
| 3 | Gemini | Google | 69.6 | overall factual claim rate % | Gemini DR (Gemini 2.5 Pro), Firm A-structured reports | Reported by the authors | Paper Table 7 | https://arxiv.org/html/2604.21006v1 | 22 Apr 2026 |
| 4 | Grok | xAI | 53.2 | overall factual claim rate % | Grok DR (Grok 4), Firm A-structured reports | Reported by the authors | Paper Table 7 | https://arxiv.org/html/2604.21006v1 | 22 Apr 2026 |

## Deep FinResearch Bench: qualitative overall (1-4, GPT-5 judge) (higher is better, 4 models)

| Rank | Model | Maker | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Gemini | Google | 2.3 | qualitative overall (1-4, GPT-5 judge) |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2604.21006v1 | 22 Apr 2026 |
| 2 | Perplexity | Perplexity | 2.1 | qualitative overall (1-4, GPT-5 judge) |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2604.21006v1 | 22 Apr 2026 |
| 3 | Grok | xAI | 2.0 | qualitative overall (1-4, GPT-5 judge) |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2604.21006v1 | 22 Apr 2026 |
| 3 | OpenAI | OpenAI | 2.0 | qualitative overall (1-4, GPT-5 judge) |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2604.21006v1 | 22 Apr 2026 |

Scores are never combined across benchmarks. Each row names its metric, source and date.

## About this benchmark

- Maker: AI Research, JPMorganChase
- Year: 2026
- Use case: Analyst Financial research
- Size: 100 professional reports covering 25 S&P 500 companies (FY2025 Q1-Q2), each paired with an AI-generated report
- Public split: none (the paper links no data or code; the professional reports came from a paid Yahoo Finance subscription)
- Grading: Automated. GPT-5 extracts claims, GPT-4.1 checks each against the cited page, and each claim ends up factual, hallucinated or non-verifiable. A GPT-5 judge gives 1-4 rubric scores, and forecasts are scored on SMAPE, hit rate and directional accuracy.
- Licence: Not stated
- Submission: none
- Products allowed: Yes
- Latest version: arXiv v1 (2026-04-22)
- What it does not cover: No public data or harness. Covers only 25 large-cap US names in IT, financials and health care, over two quarters.
- Last checked: 1 Nov 2026
- Links: paper https://arxiv.org/abs/2604.21006
- Directory page: https://benchmarks.finance/benchmarks/deep-finresearch-bench/
- Leaderboard: https://benchmarks.finance/leaderboards/deep-finresearch-bench/

Source: https://benchmarks.finance/leaderboards/deep-finresearch-bench/

Created and maintained by Moneda.