ChatGPT, Gemini, Perplexity or Grok deep research for stock reports? A bank's AI lab graded them against analysts
Every major chatbot now has a 'deep research' button that writes a multi-page report in minutes. A team at JPMorganChase's AI Research group asked four of them to write equity research on 25 S&P 500 companies and scored the output against reports from a professional research firm.
A deep research agent decomposes your question, searches dozens of pages, and writes a report with citations. For an investor, that looks like a free sell-side note. The question is how close it gets. The best published attempt to measure that for equity research is Deep FinResearch Bench, an arXiv preprint posted April 22, 2026 by nine researchers from JPMorganChase's AI Research group. It is not peer reviewed, and the paper states it is not a product of JPMorgan's research department.
How the test worked
The authors randomly sampled 25 S&P 500 companies in technology, financials and health care, took pre-earnings research reports from two professional firms (named only as Firm A and Firm B) for fiscal 2025 Q1 and Q2, and had four deep research agents write matching reports for the same company, quarter and approximate date: OpenAI Deep Research, Gemini Deep Research, Perplexity Deep Research and Grok. Reports were scored three ways: qualitative quality by LLM judges checked against human experts, forecast accuracy against reported results, and claim-level verifiability. The agents were the 2025 versions available at the time.
Quality: 'fair', with one surprise
| Author | Assumptions | Depth | Overall |
|---|---|---|---|
| Firm A (human analysts) | 1.91 | 3.73 | 2.84 |
| Gemini Deep Research | 2.00 | 3.00 | 2.31 |
| Perplexity Deep Research | 1.90 | 2.40 | 2.12 |
| OpenAI Deep Research | 1.45 | 2.60 | 2.02 |
| Grok | 1.70 | 2.60 | 2.02 |
The agents land in the 'fair' range overall. The surprise is assumptions: Gemini scored higher than the human firm on stating and justifying its modeling premises. Analysts kept a clear lead on depth and coherence.
Forecasts: closer than you might think, not better
On forecast error, measured as symmetric mean absolute percentage error (SMAPE, lower is better), Firm A had the lowest overall error at 17.14%, with Grok close behind at 17.49%, OpenAI at 21.52%, Gemini at 23.05% and Perplexity at 27.56% Context: n=25 S&P 500 companies, fiscal 2025 Q1 and Q2 reports · agents run in 2025 · Deep FinResearch Bench, arXiv:2604.21006 · margin: not disclosed, Table 4. The gap was widest on revenue (Firm A 10.64% versus 18.65% to 24.73% for the agents). Free cash flow was the exception: the human firm's error there was 51.93%, worse than every agent. The authors conclude that analysts remain more consistent across financial statement forecasting Context: n=25 S&P 500 companies, fiscal 2025 Q1 and Q2 reports · agents run in 2025 · Deep FinResearch Bench, arXiv:2604.21006 · margin: not disclosed, Table 4.
Verifiability: the number that should change how you use them
The authors extracted every claim from each AI report and checked whether the cited source supported it.
| Agent | Claims backed by cited source | Hallucinated | Could not verify |
|---|---|---|---|
| OpenAI Deep Research | 86.0% | 11.2% | 2.9% |
| Perplexity Deep Research | 75.6% | 18.5% | 5.9% |
| Gemini Deep Research | 69.6% | 17.3% | 13.1% |
| Grok | 53.2% | 34.1% | 12.7% |
Two patterns matter for traders. Numeric claims were less reliable than descriptive ones for every agent (OpenAI 84.9% versus 89.4% Context: n=25 S&P 500 companies, fiscal 2025 Q1 and Q2 reports · agents run in 2025 · Deep FinResearch Bench, arXiv:2604.21006 · margin: not disclosed, Table 7), and the numbers are what you would act on. And the best qualitative writer, Gemini, was not the most factual; the paper finds Gemini and Grok cited a significantly lower share of high-trust sources (its high-trust examples include sec.gov, reuters.com and company IR sites; its low-trust examples include x.com, investing.com, macrotrends.net and seekingalpha.com).
What access costs (read October 14, 2026)
- ChatGPT: deep research is 'limited' on Free and Go, 'expanded' on Plus and 'maximum' on Pro, per OpenAI's pricing page; dollar amounts did not render in the version we opened.
- Claude: its Research feature is not on Free; Pro is $20 a month, or $17 a month billed annually (claude.com/pricing). Claude was not in this benchmark.
- Gemini: Deep Research is listed from the Google AI Plus plan up (Google One); prices did not render.
- Perplexity and Grok: we did not obtain current plan pages. For how the newest models score on public finance benchmarks, see our AI Model Index.
Using a deep research report as homework, not as research
- Treat every number as a lead. Even the best agent had about one claim in nine judged hallucinated.
- Click the citation, not the summary. Low-trust sources were a measurable failure mode.
- Ask for assumptions explicitly. That was where agents matched or beat the human firm, and it makes the report checkable.
- Compare forecasts with consensus yourself. Agents were furthest off on revenue, the line most models start from.
What it shows / What it doesn't
It shows
- A structured comparison of four deep research agents with professional equity reports on the same 25 companies and quarters.
- That agents produced 'fair' reports, matched analysts on stated assumptions, and trailed on depth and most forecasts.
- Claim-level hallucination rates between 11.2% and 34.1% across the four agents Context: n=25 S&P 500 companies, fiscal 2025 Q1 and Q2 reports · agents run in 2025 · Deep FinResearch Bench, arXiv:2604.21006 · margin: not disclosed, Table 7.
It doesn't show
- Current versions. The agents tested were 2025 releases, and all four products have changed since.
- Claude or other research agents, which were not in the benchmark.
- Independent replication. It is a preprint from one bank's AI lab, and the professional firms are anonymized.
- Whether acting on an AI report makes money. Stock-level results were reported by horizon and varied between agents.
Why it matters
A deep research agent can assemble in minutes the background a human analyst spends days on. This test shows its numbers need the same checking as any unsourced note, and that the call on the stock remains yours.
Primary sources
- VerifiedDeep FinResearch Bench: Evaluating AI's Ability to Conduct Professional Financial Investment Research (arXiv:2604.21006) (arXiv, JPMorganChase AI Research, April 22, 2026)Abstract and authors read October 14, 2026.
- VerifiedFull text, HTML version (arXiv)Tables 3, 4 and 7 and method sections used for every figure above.
- VerifiedChatGPT plans and pricing (OpenAI)Deep research access by plan read October 14, 2026.
- VerifiedClaude plans and pricing (Anthropic)Research availability and Pro price read October 14, 2026.
- VerifiedGoogle AI plans (Google One)Deep Research listed on AI Plus, Pro and Ultra, read October 14, 2026.
In our directories
How we used AI: AI tools helped research and draft this article. Every figure was then checked against the primary sources listed above before publication.
Not financial advice. For information and education only. AI model outputs and research results can be wrong and do not predict future returns. Nothing here is a recommendation to buy or sell any asset. See our disclaimer and methodology.