OpenAI just released GPT-5.2 yesterday. This is their “code red” response to Google’s Gemini 3.
Buried in the release notes is something finance professionals should care about.
OpenAI benchmarked GPT-5.2 on junior IB analyst spreadsheet tasks (three-statement models, LBOs, formatting).
Result: 68.4% accuracy vs 59.1% for GPT-5.1.
A 9.3 point jump in one release.
Other upgrades that matter for finance:
- Spreadsheet formatting and financial modeling explicitly improved in Thinking mode
- 196K context window on paid tiers (enough for several 10Ks in one chat)
- <1% hallucination rate on financial/tax domains with browsing enabled
- 30% fewer errors overall vs GPT-5.1
This opens up a use case that was previously unreliable: building a full DCF model directly in ChatGPT.
Previously, AI would fumble the math.
Discount factors wrong.
Terminal value hallucinated.
WACC inconsistent with projections.
More time debugging than just building it yourself.
I wanted to test if GPT-5.2 actually delivers.
So I built a complete DCF for Uber ($UBER) using 2 prompts on GPT-5.2 Thinking.
The plan:
- Upload Uber’s latest financials
- Paste in our pre-prepared prompt
- AI outputs a downloadable Excel file with sensitivity tables
This takes a junior analyst 2-3 hours. We’re doing it in minutes.
The Prompt
Make sure you use "Thinking" mode in the new ChatGPT 5.2 model
I also highly recommend downloading at least the last 5 quarters worth of 10Ks and 10Qs to upload to the prompt. We need to ensure data fidelity for complex prompts like this. Here's a free site I use: Link
We will be using 2 prompts to produce this DCF (to make sure the model doesn't brick while calculating too many things).
The first prompt calculates all the assumptions using the 10Ks and 10Qs.
The second prompt takes these assumptions and builds the actual Excel file.
Prompt 1: DCF_Prompt_Template_Chatgpt_5.2_prompt1.pdf
Prompt 2: DCF_Prompt_Template_Chatgpt_5.2_prompt2.pdf
The Result
Full output here: Link
Download the DCF Output: UBER_Quarterly_Unlevered_DCF.xlsx
I was incredibly impressed with the output. Although this was quite slow (took ~45 minutes) the quality was very very solid.
This is the type of work that looks like it came from the desk at Goldman.
Is this perfect? No.
But we just saved hours and hours of time to get the model to 90% of the finalized structure. It is now your job to stress test these assumptions and bring in the high value work of "thinking".
In just 6 months, we have gone from "AI in Excel is totally useless" to "AI can replace bottom to mid bucket analysts easily".
(ChatGPT also just killed an entire cohort of VC backed finance excel companies)
I really do expect to have perfect models being built using ChatGPT or one of the frontier models in the next 12 months .... amazing times we're living in.