Key Takeaways

  • OpenAI does not train isolated, proprietary models for finance; it feeds financial reasoning data directly into the core foundation model run, dubbed "the big run" (such as runs like 6 Astra or 61 Astra).
  • Because financial intelligence sits in the main model, external developers get the exact same underlying financial reasoning capabilities as OpenAI's internal product teams.
  • Joseph Kim built 95 percent of ChatGPT for Financial Services using the exact same ChatGPT and Codex plugin primitives publicly available to external startups.
  • Base foundation models and Codex get vertical software roughly 80 percent of the way to market readiness, leaving the remaining 20 percent as the territory where vertical startups win.

OpenAI Trains Financial Data Directly Upstream

When founders build vertical AI tools for finance, they often assume OpenAI hoards proprietary financial models behind locked doors. Joseph Kim, who leads ChatGPT for Financial Services at OpenAI, says the internal setup is the exact opposite. OpenAI feeds domain-specific financial training directly into the primary base model.

“All of the training data that we do for financial, you know, reasoning analysis, we do upstream into what we call the big run, which is the the main kind of run that becomes, you know, something like 6 Astra or 61 Astra, etc.,” Kim explains.

This training choice has immediate consequences for external software builders. “Even though we're adding a bunch of financial data to the model, it's going into the main main model, and so therefore it's available to folks that want to build on top of it,” Kim notes. If you build financial analysis software, you do not need to train your own foundational finance model from scratch. OpenAI does that heavy lifting in its core runs, democratizing the raw reasoning power for every developer via standard APIs.

The 80 Percent Baseline and the 20 Percent Wedge

If the base foundation model already contains deep financial reasoning, what prevents OpenAI from swallowing vertical finance software whole? Kim points out that horizontal foundation models hit an architectural ceiling when applied to production finance workflows.

“The Codex is fantastic,” Kim says. “It's got great memory. It's got good skills. Like there's a you know, it's persistent. There's a lot of things that are wonderful about Codex, but the thesis of my team was Codex will get you about 80% of the way to a product that would be successful in the market, but the 20% is missing.”

Surprisingly, OpenAI's internal finance team did not rely on exotic, unreleased tools to close that gap. Kim reveals that “95% of what we built was something that absolutely a startup could do, another company could do.” They assembled the product using basic Codex and ChatGPT plugin primitives. They solved the remaining 20 percent by bundling reliable financial data, replacing flaky third-party connectors, and designing specific human review chains.

That missing 20 percent is where vertical software companies survive. Horizontal models cannot guess the specific regulatory compliance paths, custom ledger integrations, or edge-case review loops required by enterprise finance desks. “I really do think that there's going to be a billion-dollar business built on top of a ChatGPT plugin,” Kim states. “You're going to be able to leverage all of that horizontal functionality and just add your secret sauce on top.”

What to Do With This

Stop spending capital attempting to pre-train or fine-tune foundational financial reasoning models. Instead, inspect your product's architecture this week and list every workflow where your system relies on third-party API connectors for market or balance sheet data. Replace unverified data streams with direct, bundled data sources, and spend your engineering sprints building the human review verification chain that base foundation models cannot deliver on their own.