I Switched from Claude’s Sonnet to OpenAI’s O1 Model for Financial Research. Finance Jobs WILL NOT Survive.

With O1, I could ask the challenging financial questions shown here and inspect its first-attempt answers within minutes. That convenience was why I preferred it for these tasks.

For example, I asked:

what about these two groups:
* all AI stocks
* all AI stocks excluding the stocks in the top 10 percentile and bottom 90 percentile of returns

This time, I want to weight it by market cap, so higher market caps as of last year are weighted more than lower market caps

It returned the filtered analysis on the first try. The retained group still needs the predicate and row checks described below.

The reported outputs are about 45% market-cap-weighted return for all AI stocks and 33% for the filtered group. The request above says to exclude the top 10 percentile and bottom 90 percentile, which does not define a usable retained middle band. Before accepting the filtered figure, read Appendix 3’s exact predicates and the linked returned rows, state the retained percentile band and stock count, and verify the weights against the prior-year market-cap date. Those details are not given in this prose.

I can’t reiterate enough – I retrieved this answer in minutes. Even an experienced, highly intelligent SQL expert would likely need 30 minutes to several hours to write the same query.

I am dumbfounded. I couldn’t even perform the same analysis on Claude. I would need to do some complex prompt engineering, advanced techniques like retrieval-augmented-generation, and god-forbid fine-tuning.

But with O1, I can just ask my question.

In these displayed examples, o1 returned useful-looking queries without the additional prompt work I had needed with Claude. That is my January 2025 observation. Inspect the source dates, selected universe, grouping and arithmetic before accepting it. Read the original conversation and queries.

Concluding Thoughts

For some reason, we moved the goalpost on what artificial super intelligence means.

If we use our common sense, we know that it means an AI model that's smarter than smart human beings.

And I don't know about you, but I personally know maybe one or two people that is smart enough to write a database query complex as OpenAI’s O1 model.

I preferred this recorded O1 response to the shown Claude 3.5 Sonnet response because it gave me a useful first query to inspect. The filtered result still needs the predicate, period and source-row checks described here.

Now, I'm not going to claim that this is perfect. Because of the significant increase in accuracy, we might be more prone to trusting the results out of the box whereas we may have applied more scrutiny with other models.

That is objectively a bad thing — the costs for mistakes for this model are higher because we hold it to a higher standard.

But objectively, this is a good problem to have. The OpenAI O1 model is super intelligent, but every reasonable definition of the words.

And this is the dumbest it’ll ever be.

My prediction is that this will transform traditional finance jobs over the next decade. What I demonstrated here is historical query generation and result formatting. Defining a sensible question, checking its data and interpreting what the answer means remain separate work; this comparison does not measure an entire analyst’s job.

For the query-writing work shown here, I expect models to change the time and staffing needed. The article does not provide a timed human comparison or a full-workflow cost and accuracy study.

The world will be a much scarier place in 2026.

Thank you for reading! As of January 5th, 2025, there are VERY few platforms that allow you to utilize OpenAI O1 for financial research. In my research, I found literally one: my app NexusTrade.

Prepare a defined weighted-return research request. Review the universe, endpoint dates, retained percentile band, query and stock counts before accepting an average.

Thousands of already have. Why haven’t you?

Appendix

The following appendix shows the exact queries and thought processes generated by the model. These are important for verifying the claims I make in this article, but wouldn’t flow well to insert into the middle of the article.

1. Claude 3.5 Sonnet query for “what is the average return of AI stocks in the past year”

2. OpenAI O1 query for “what is the average return of AI stocks in the past year”

3. OpenAI O1 query for “what about these two groups: all AI stocks and all AI stocks excluding outliers, with returns weighted by market cap”