Claude vs ChatGPT for HK Financial Analysis: 4 Tests I Ran
Contents
Claude vs ChatGPT for HK Financial Analysis: 4 Tests I Ran
Last Updated: 2026-05-26
By Jim Liu — 27 Hong Kong IPO subscriptions tracked, HK/AU retail investor, maintainer of the LRTS 107-IPO dataset.
TL;DR
- I gave Claude (Sonnet 4.5) and ChatGPT (GPT-4o) the same four prompts I actually run as part of my weekly HK analyst workflow.
- Claude won the long-context tasks — HK IPO prospectus parsing and HKMA stablecoin regulatory reading — because it handled 200+ page PDFs in one shot without losing thread.
- ChatGPT was the better calculator partner for dividend DCFs and multi-broker fee comparisons, especially when I needed it to call a code interpreter.
- Neither is a licensed advisor. Both hallucinated HK-specific regulation when I pushed them past their cutoff. The audit trail matters more than the headline answer.
- Cost for a full analyst week: roughly USD 25–40/mo if you pick one, around USD 60 if you run both like I do.
I started running Claude and ChatGPT side-by-side about ten months ago, after a futu prospectus chewed up most of a Sunday. The question I kept getting from readers — and from my own newsletter inbox — is which one is actually better for Hong Kong financial analysis. Not "which is better at writing essays." Specifically: which one would I trust to help me read a Chapter 18C tech prospectus, model a dividend stock, compare broker fees, or parse the HKMA's stablecoin framework.
This is the honest answer, four tests, one analyst, no affiliate angle. I pay for both subscriptions myself.
A note before we start. I'm a retail investor, not a licensed advisor, and not an AI researcher. I'm running both tools as a daily analyst user, on real work, and reporting what came out. Your mileage absolutely varies.
Why I Tested Both: The HK Analyst Context
There's a specific kind of work most retail-facing HK analyst content skips, and it's the work I actually do most weeks.
It's three buckets, roughly. Long-document parsing — HK IPO prospectuses run 300–600 pages in PDF, often with the English and Traditional Chinese versions stitched together. Quantitative modelling — dividend yield projections, broker fee tables with three-decimal HKD precision, FX-adjusted DCFs. Regulatory reading — HKMA circulars, SFC guidelines, the new stablecoin ordinance, all of which moved this year.
When the AI tools first arrived I assumed they'd all be roughly equivalent for this. Ten months in, they aren't. Each has a personality, and the personality matters when you're trying to decide whether to commit capital on an IPO ballot day.
So I ran four tasks. Same prompt to both. Same data input. I scored on Hong Kong regulatory accuracy, calculation reliability, audit trail (can I see how it got there), and cost. The full per-use-case verdict is below.
Test 1: HK IPO Prospectus Analysis
The first test was the messy one. I picked a recent Chapter 18C tech listing — one of the autonomous-driving names from this year's queue — and uploaded the English prospectus PDF, around 412 pages.
My actual prompt was: "This is the Chapter 18C prospectus for [redacted]. I'm a retail investor considering subscribing. Pull out (1) the pre-money valuation and how that compares to the last private round, (2) cornerstone investor list with lock-up periods, (3) any unusual risk-factor language, (4) cash burn rate and runway. Quote page numbers."
ChatGPT (GPT-4o, file upload): It got the cornerstone list right and was fast — answer in about 90 seconds. The valuation read was off by one funding round (it mixed the Series D with the Series E, which would have been a meaningful misread if I hadn't cross-checked). Page citations were present for about half the claims; the other half it just stated as fact.
Claude (Sonnet 4.5, file upload): Slower, around three minutes for the first pass. But the valuation reconciliation was correct, and when I asked it to compare cornerstone lock-ups to a comparable 2025 listing it actually re-read the relevant pages instead of guessing. The risk-factor pull was the part that surprised me — it flagged a clause about "ongoing PRC regulatory review of L4 autonomy permits" buried in the legal disclosures that I had genuinely missed on my own read-through. That clause turned out to matter on listing day.
I went back and forth on this one for two weeks across three different prospectuses to make sure it wasn't a fluke. The pattern held. Claude was more accurate on long PDFs, especially on the legal language; ChatGPT was faster but I caught it cutting corners on citations.
For context, I cover the actual prospectus-reading process in my Hong Kong IPO 2026 retail field guide, and the Chapter 18C rules specifically in the 18C specialist tech IPO guide. Both pieces inform what I'm asking the AI to actually find.
My verdict for this use case: Claude. Not by a small margin either.
Test 2: Dividend Stock DCF
This was the test I expected ChatGPT to win, and it did.
I asked both to build a discounted cash flow on a HK blue chip dividend name — one of the names I tracked through the LRTS blue-chip dividend yield rankings. I supplied the last five years of dividend history (which I pulled from my own data, not theirs), my own assumed discount rate (8.5%), and a terminal-growth assumption (2.0%). I asked for a per-share fair value plus a sensitivity table varying terminal growth from 1.5% to 3.0%.
ChatGPT (GPT-4o, code interpreter on): Wrote the DCF in Python, ran it, returned a fair value of HKD 78.40 and a clean 6-by-6 sensitivity table. I could see the code, re-run it, modify the assumptions. About 40 seconds.
Claude (Sonnet 4.5): Did the math in its head, basically. Walked through the calculation in prose, returned a fair value of HKD 79.10. The sensitivity table was a markdown table rather than a re-runnable artifact. Took longer to type out, around two minutes.
Both got within HKD 1 of each other, which is reassuring. But the audit trail mattered more than the answer. With ChatGPT I could change the discount rate to 9.0% and re-run; with Claude I had to re-prompt and trust it. For modelling work where I'm going to vary assumptions five or six times, the code-interpreter loop is the workflow that wins.
Caveat: I cross-checked both with my own Excel model. Both got the math right at the headline level. The cases where AI quietly fails are usually in tax assumptions — HK dividend tax is 0% domestic, but US stock dividends held by HK residents trigger 30% IRS withholding (or 10% with W-8BEN routing through certain brokers), and I've seen both tools occasionally forget to apply it. I've written about that in HK stock dividend tax withholding and the US stock dividend tax for HK investors.
My verdict for this use case: ChatGPT, on workflow grounds, not accuracy grounds.
Test 3: Broker Fee Comparison Reasoning
This one was supposed to be the easy test. It turned out to be the test that exposed how both tools handle ambiguous criteria.
I asked both: "I'm a HK retail investor planning around HKD 500,000 in annual turnover across HK stocks, US stocks, and HK IPO subscriptions including margin. Compare Futu, moomoo, Tiger Brokers HK, IBKR, and HSBC SecuritiesDirect across (a) per-trade cost, (b) HK IPO margin rate, (c) FX spread on USD conversion, (d) custodian fee handling. Recommend one."
ChatGPT: Returned a recommendation in around 60 seconds (Futu, with moomoo as the backup). The per-trade fee numbers were close to current but slightly outdated by what looked like six months. The margin rate figure for one broker was just wrong — it quoted a promo rate as the standard rate.
Claude: Took longer, around two minutes. Refused to give a single recommendation without first asking three clarifying questions: how many of the HKD 500k would be IPOs vs secondary trading, was the user comfortable with overseas custodian setup, and did the user care about ETF coverage. After I answered, the recommendation came back as Futu or IBKR depending on the IPO/secondary mix, with explicit reasoning. Per-trade fee numbers were also slightly stale but on roughly the same vintage.
Neither tool was current enough to trust on raw fee numbers — I cross-check those against my own HK stock broker comparison and the best broker for HK IPO beginners writeups, both of which I update from primary sources.
The difference between the two AIs here was style. ChatGPT gave me one answer fast; Claude made me think about the problem properly before giving an answer. Whether that's good or annoying depends on what you're doing. For a quick gut check, ChatGPT's pace is fine. For something I'm about to commit capital to, I'd rather be slowed down by Claude's clarifying questions.
My verdict for this use case: Claude, mostly because of the clarifying-questions habit. I count that as accuracy, not annoyance.
Test 4: HKMA Stablecoin Regulatory Parsing
This was the test where I cared most about hallucination risk.
The HKMA's stablecoin ordinance went into effect this year, and the licensing framework matters for anyone touching tokenized HKD products. I gave both tools the official HKMA consultation paper (~80 pages, English) and asked: "What's the minimum capital requirement for a licensed stablecoin issuer? What's the reserve composition rule? Are there fiat-denominated stablecoins exempt from the regime? Quote the section."
ChatGPT: Answered confidently. HKD 25 million minimum capital — correct. 100% reserve in high-quality liquid assets — correct in spirit but it mis-stated the eligible asset list (it included corporate bonds, which the regime does not include for the reserve pool, only HKD-denominated short-dated government and central bank assets). It did not quote a section number; when I asked it to, it provided one that didn't match the document.
Claude: Slower, more cautious. HKD 25 million confirmed with the section quoted accurately. The reserve composition answer was correct and conservative — it explicitly said "I see HKD-denominated short-dated sovereign and central bank assets in the eligible list; corporate bonds do not appear in the section as drafted." On the exemption question, it flagged ambiguity rather than inventing certainty.
This was the test that decided how I use these tools going forward. For HK regulatory work specifically, Claude was the lower hallucination risk. ChatGPT's confident-but-wrong eligible-asset list would have caused real harm if I'd quoted it in a piece for readers.
I keep the HKMA stablecoin regulation guide as my own reference, and I now cross-check every AI summary against it before publishing.
The Verdict, Per Use Case
This is the table I actually use to decide which tool to open for which task.
| Use Case | Claude (Sonnet 4.5) | ChatGPT (GPT-4o) | Better for HK Analyst |
|---|---|---|---|
| HK IPO prospectus (300+ page PDF) | Higher accuracy on legal language; cites pages reliably | Faster but cut citation corners; mixed funding rounds | Claude |
| Dividend DCF / re-runnable model | Prose math, no re-runnable artifact | Python code interpreter; iterate assumptions cleanly | ChatGPT |
| Broker fee comparison reasoning | Asks clarifying questions before recommending | Fast single recommendation; quoted a stale promo rate as standard | Claude |
| HKMA stablecoin / regulatory parsing | Conservative; flags ambiguity; section quotes match | Confident; eligible-asset list was incorrect; fabricated a section number | Claude |
| Calculation reliability | Good on simple math, no code execution | Excellent when code interpreter is on | ChatGPT |
| Audit trail (can you see the reasoning) | Strong — quotes source, shows reasoning steps | Mixed — code is great, prose reasoning is sometimes shallow | Claude |
| Cost (analyst plan) | USD 20/mo (Claude Pro) | USD 20/mo (ChatGPT Plus) | Tie |
A genuine downside on Claude: there's no code interpreter inside the chat product the way ChatGPT has one. For repeated model iteration, that's a real workflow gap. You can paste outputs into your own Python environment, but it's not the same as the back-and-forth.
A genuine downside on ChatGPT: it will state things with confidence when it shouldn't, and the section-number fabrication on the HKMA test wasn't a one-off — I've seen the same pattern on SFC circular questions.
Cost Analysis: What an Analyst Workflow Actually Costs
If you pick one, you're roughly looking at USD 20/month — Claude Pro or ChatGPT Plus, depending on which way you went. That covers maybe 80% of the analyst use cases either tool handles well.
If you run both, like I do, the bill comes in around USD 40/month, plus a small API tab if you're scripting batch jobs against the developer endpoints (mine is maybe USD 10–20 in a heavy month, mostly Claude API for prospectus summarization in bulk).
For a single retail investor running personal portfolio work, the dual-subscription cost looks expensive next to "free Google search." For someone actually publishing analyst-quality work or making capital decisions across multiple positions a week, it's trivially cheap compared to one wrong call on a HKD 500k position. That's the framing I use.
There's no equivalent affiliate funnel here. I pay for both. There are cheaper open-source alternatives (Llama 3, DeepSeek) for parts of the workflow, but the regulatory-language accuracy gap on closed-domain HK content is real, and I haven't found an open-weights model that closes it for my specific use cases yet.
FAQ
Is Claude or ChatGPT better for reading a Hong Kong IPO prospectus?
In my testing across three different prospectuses, Claude was meaningfully more accurate on long-PDF legal language and citation reliability. ChatGPT was faster but more likely to mix funding rounds or skip page citations.
Can I use ChatGPT for HK dividend stock DCF modelling?
Yes, and this is the use case where ChatGPT's code interpreter is genuinely better. You can iterate discount rate and terminal growth assumptions in a re-runnable Python artifact, which Claude doesn't offer inside the chat product. Verify the math against your own model regardless of which tool you use.
Did either tool hallucinate on HKMA stablecoin regulation?
ChatGPT stated an incorrect eligible-asset list and fabricated a section number when pressed for a citation. Claude was more conservative, flagged ambiguity where it existed, and quoted sections that matched the source document. For HK regulatory work specifically, the hallucination risk asymmetry is real.
What does an analyst-grade AI workflow cost per month?
Around USD 20/month for one (Claude Pro or ChatGPT Plus), around USD 40/month for both. Adding API usage for batch jobs typically pushes a heavy month into the USD 60–80 range. Free-tier versions of both are insufficient for real analyst document work because of context-length and rate limits.
Should I use AI to make actual investment decisions?
No, and neither tool will tell you to. Both are useful as research and reading accelerants. Both will occasionally get HK-specific regulatory or tax detail wrong. Final decisions on capital allocation still need to come from primary sources and, where appropriate, a licensed advisor. I treat AI output as a first pass to be cross-checked, never as the answer.
Why did you not test Gemini or other models?
I do use Gemini for image and chart extraction, but for the specific use cases here — long PDF prose, regulatory language, financial modelling — Claude and ChatGPT are the two I run weekly. I'll cover Gemini in a separate piece once I've put it through the same four tests.
Methodology Note
Each test was run on the same day, on both tools, with the same input. Where file upload was used, the same PDF was used. Where math was involved, I cross-checked against an Excel model I maintain for my own portfolio work. I tested Claude Sonnet 4.5 (via Claude Pro) and ChatGPT GPT-4o (via ChatGPT Plus), versions current as of late May 2026. Model versions change frequently; the personality differences I describe may shift with each release. I'll update this piece when the verdict changes materially.
Not financial advice. This is an analyst's working notes on AI tools, not investment guidance. AI tools occasionally hallucinate on Hong Kong regulatory, tax, and securities detail — including, in my testing, both Claude and ChatGPT. Consult a licensed advisor for HK securities decisions and verify all regulatory claims against primary sources (HKEX, HKMA, SFC). Past performance of any IPO subscription, dividend stock, or broker comparison is not indicative of future results.
About the author: Jim Liu is a retail investor based between Hong Kong and Australia. He maintains LowRiskTradeSmart (LRTS) and tracks Hong Kong IPO market data, including the 107-IPO historical dataset on this site. He has personally subscribed to 27 Hong Kong IPOs since January 2024 across cash and margin routes.