OpenAI Built a Product for the Work Junior Bankers Do

For two generations, the way you learned investment banking was by doing the worst job in it: pulling filings past midnight, reconciling an adjusted EBITDA figure against the footnote explaining which costs management excluded, dropping the result into PowerPoint, then walking the deck backward slide by slide to confirm every number traces to something real. On Thursday, OpenAI shipped a product built to do that in minutes.

ChatGPT for Financial Services, announced September 10, 2026, is a tailored version of ChatGPT Work running on GPT-6 Astra, the frontier model OpenAI released a day earlier. It was developed with Morgan Stanley and Evercore as design partners, and it launches aimed squarely at the two functions where the grind is heaviest: investment banking and equity research.

OpenAI is not being coy about the target. “We're effectively teaching ChatGPT to research like an analyst and back up its conclusions like an analyst as well,” said Nick Turley, OpenAI's head of ChatGPT, at a press roundtable accompanying the launch.

What distinguishes the product from a general chatbot with a finance prompt is plumbing. Premium datasets from Daloopa, PitchBook and LSEG News, covering earnings transcripts, financial statements, company fundamentals and private-company coverage, come bundled in, indexed and hosted on OpenAI's own infrastructure rather than fetched from a vendor at query time. No separate data contracts, no connectors to wire up. A second path lets firms carry over subscriptions they already pay for: OpenAI says it is building shared sign-in and entitlement integrations with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's, so the model inherits a user's existing licences instead of becoming a route around them. A third path is the connector ecosystem, now more than 50 integrations. OpenAI says tuning cut connector error rates from 7.53 percent to 2.57 percent for Daloopa and from 6.84 percent to 2.66 percent for S&P Global.

The output side is where the junior-banker comparison gets literal. Administrators publish approved Excel, Word and PowerPoint templates through a dedicated admin page, and the system renders valuation models, research notes and pitchbooks in the firm's own house format. Every figure carries granular citations highlighting the table or passage it came from. In a live demo, Turley had the product evaluate a prospective acquisition target, pull financials from industry-standard feeds and generate a bank-templated deck. In another, it turned model outputs into an interactive web page where assumptions could be changed and flowed through the projections.

OpenAI reported GPT-6 Astra scoring 69.9 percent on OfficeQA Pro, a test of 133 questions across roughly 89,000 pages of U.S. Treasury Bulletins, against 60.2 percent for GPT-5.6 Sol and 62.4 percent for Claude Fable 5.1. Turley was dismissive of leaderboards anyway. “Academic benchmarks are not interesting to this industry,” he said. “It is about real-world utility, and that's on cost and on utility.” He claimed Astra is roughly twice as efficient as the strongest alternative measured by cost per task. OpenAI disclosed no pricing, minimum seat counts, geographic availability or eligibility criteria, and declined to name banks beyond the two design partners.

Asked whether the product would shrink junior banker hiring, Turley reached for the obvious historical analogy. “If you study the life of an analyst or of a banker, depending on the industry, they're working 100-hour weeks,” he said. “I think in the same way that Microsoft Excel transformed the industry and allowed them to produce better analysis faster, you will see technology like this do the same.”

The Pipeline Is the Product Being Disrupted

The Excel comparison is the most load-bearing claim in the launch. Excel expanded what one analyst could produce; it did not attempt the analyst's judgment or write the memo. This product explicitly targets the full artifact, from question to client-ready deck. That matters because the tasks being automated are not merely tedious. They are the apprenticeship. An associate learns which adjustments are defensible by being made to defend them at two in the morning.

Wall Street is already cutting into that pipeline. Reporting this year has described Goldman Sachs, JPMorgan, Citi and Barclays trimming junior analyst classes by as much as two-thirds as AI absorbs entry-level work, with Goldman president John Waldron describing parts of bank operations as a human assembly line ripe for automation. UBS has taken the other tack, telling candidates that from the 2027 intake AI fluency will sit alongside academics and finance aptitude as an explicit hiring criterion, with interviewers probing how applicants actually use the tools, including whether they know when the model is wrong.

That last clause is the real stake. In banking, a hallucinated number is not an embarrassing chat transcript; it is a fairness opinion, a regulatory filing, a valuation a client acts on. OpenAI's answer is provenance rather than confidence: trace the claim, surface the table, let the professional judge. It is the right architecture, and it also quietly relocates the verification burden onto whoever is left to verify. If the analyst class that historically did that checking is half its former size, the checking does not disappear. It just gets thinner.

For vertical AI more broadly, this is the template. The model is increasingly table stakes; the moat is bundled licensed data, inherited entitlements, firm templates, information barriers and exportable audit logs. OpenAI is competing for the interface layer, and Turley said so plainly: “this is the canonical product that we hope the industry adopts.” Anthropic reached finance first with Claude for Financial Services, Microsoft is wiring the same feeds into Copilot, and FactSet and S&P Global are building upward from their own data.

Watch three things next. Whether any bulge-bracket firm beyond the design partners commits publicly, and on what pricing. Whether OpenAI's promised sector-tailored versions for other industries arrive on the same bundled-data model. And whether the 2027 analyst classes at the banks deploying this shrink, hold, or get quietly rebuilt around people hired to supervise the machine rather than to be it.