OpenAI's finance ChatGPT ships with two design partners, no named customer and no price
The product bundles Daloopa, PitchBook, LSEG and Crunchbase data on OpenAI's own servers, and proves itself on a benchmark made of public Treasury bulletins.
OpenAI's new product for banks arrived with two design partners, no customer willing to be quoted, and no price anywhere on the page.
ChatGPT for Financial Services launched on 10 September 2026. It is ChatGPT Work with earnings transcripts, company fundamentals and private-market records from Daloopa, PitchBook, LSEG News and Crunchbase already indexed and hosted on OpenAI's own infrastructure. Morgan Stanley and Evercore are named as the firms that shaped it. Nobody at either is quoted saying they use it.
So what did a bank get on Thursday that it could not get from ChatGPT Enterprise on Wednesday?
Mostly the data, and the plumbing under it
The security model is the old one: SAML sign-on, automated user provisioning, role-based access controls, admin-set retention, log export. What is new is narrower and more interesting. Premium datasets ship with the product instead of through a contract you negotiate yourself. Citations are granular enough that, in OpenAI's framing, a banker can "trace figures and claims back to their sources". Administrators can publish Excel, Word and PowerPoint templates from a dedicated page, so a pitchbook comes out in the firm's own format. And a firm can run several workspaces to enforce information barriers — the compliance desk's phrase for keeping the M&A team out of the research team's chats.
Those templates are the most honest thing in the announcement. Somebody at OpenAI worked out that the job is not analysis. The job is a deck by Sunday.
The benchmark is about the wrong documents
OfficeQA Pro tests whether AI agents can find and analyze information across U.S. Treasury Bulletins, including complex financial tables, charts, and supporting footnotes. GPT-6 Astra scores 69.9%, compared with 60.2% for GPT-5.6 Sol.
Read that caption twice. The evaluation chosen to prove a finance product measures how well a model reads Treasury bulletins, which are public, free, and sitting on a government website. What customers are paying for is the Daloopa and PitchBook corpus, and there is no number anywhere for how the model handles that. A gain of nearly ten points on public PDFs is evidence about PDFs. No bank has ever bought software because it read a Treasury bulletin well.
This is a data-licensing product wearing a model announcement's clothes, and the competitor that matters is the terminal on the desk rather than anything Anthropic ships.
Anthropic did this fourteen months ago, with customers talking
Claude for Financial Services launched on 15 July 2025, and it launched with people on the record: Bridgewater's AIA Labs, Commonwealth Bank of Australia, and AIG's chief executive describing a five-fold compression in underwriting review time. OpenAI has design partners and no testimony. Two of the four providers OpenAI bundles by default, Daloopa and PitchBook, sat in Anthropic's partner list that same month, which tells you roughly how differentiated the underlying data is. Both labs are renting the same pipes.
Take the counter seriously, because it is a good one. Design partners who shape post-training are worth more than a marketing quote, and neither Morgan Stanley nor Evercore lends its name casually — investment banks are famously allergic to appearing in a vendor's launch. The entitlement work with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's is the genuinely hard part, because it means a provider can recognise you through your ChatGPT sign-in and release data you already pay for. Do that well and ChatGPT becomes the surface where a five-figure annual subscription gets consumed. That threat to the incumbents is real, and it has nothing to do with reasoning.
So watch the plumbing, not the leaderboard: whether any of those five subscription providers says the integration is live and general, or whether it stays at "we're working with". Entitlement deals die quietly. We would expect no named bank to say publicly that it has deployed this to its analysts before the end of the first quarter of 2027, and if one does, with a headcount attached, we were too sour about a pilot that turned out to be a product.
Bankers will find out sooner than that. The first analyst to put an Astra-built comp table in front of a managing director without checking the footnotes will produce the industry's real evaluation, and it will not appear on a leaderboard.
