OpenAI introduced ChatGPT for Financial Services on September 10, 2026, as a tailored ChatGPT Work workspace for research, models, and client materials. The first cut is aimed at investment banking and equity research. Design partners named in the post are Morgan Stanley and Evercore.

The pitch is data that does not depend on each bank wiring its own connectors. OpenAI said it indexes and hosts premium sets from Daloopa, PitchBook, LSEG News, and Crunchbase, covering earnings transcripts, statements, fundamentals, and private companies. Bankers can start without a separate contract for those feeds. The product adds granular citations so a figure can be traced to a table or passage. GPT-6 Astra is included, with newer models promised as they ship, under ChatGPT’s existing enterprise access and governance controls.

Workflows the banks asked for

OpenAI said the early work with Morgan Stanley and Evercore pointed to two pain points: reliable data and high-quality artifacts. The pages shown in the announcement include value analysis, leveraged buyout modeling, buyer screening, and earnings analysis. A banker running a P&L normalization, the post says, should be able to inspect the notes behind an adjusted EBITDA and decide how to use it.

Morgan Stanley said it is using the collaboration to shape how the technology develops around the firm’s own methods. Evercore said the test is whether frontier models deepen the advice without replacing the judgment clients already pay for. Jager McConnell, CEO of Crunchbase, said the hosted private-company file is meant to cut the time from a question to a usable answer.

OpenAI said it will keep post-training models on these datasets and expand coverage with the same vendors. It did not publish a seat price, a minimum, or a public signup link.

Decoded Take

Banks have been bolting models onto terminals for two years. OpenAI is trying to own the retrieval layer so the model is not guessing from the open web. Hosted PitchBook and LSEG feeds inside ChatGPT Work will scare the vendors that still sell those terminals, and it will scare compliance teams that have to explain where a citation came from. The useful test is not a polished comps file in a demo. It is whether a managing director will send a book that the model drafted without a human reconstructing every number. Watch which other banks join Morgan Stanley and Evercore, and whether the data vendors start selling the same hosted bundle to Anthropic and Google.