Every day, investment-banking analysts spend hours pulling company and market data, reconciling adjustments in Excel, checking assumptions, moving the analysis into PowerPoint and then working back through the deck to make sure every important figure can be traced to the underlying source.
Now, OpenAI wants ChatGPT to do all that work in mere minutes, even seconds — without severing the audit trail that lets the banker verify the work.
That's the reason the company today launched ChatGPT for Financial Services, a new industry-specific version of ChatGPT Work, its productivity harness and mode, that combines its new, cutting-edge AI model GPT-6 Astra with premium financial data, firm-controlled templates and enterprise governance.
OpenAI developed the product with major investment banks Morgan Stanley and Evercore, whose feedback helped shape its initial focus on investment banking and equity research. The firms reportedly highlighted reliable data access and high-quality artifact creation as two of the biggest pain points for their teams.
OpenAI is also working with a broader group of financial-data and software providers. The product includes premium data from providers such as Daloopa, PitchBook and LSEG News; OpenAI is developing shared sign-in and entitlement integrations with S&P Capital IQ, LSEG, MSCI, Factiva and Moody’s; and it is optimizing connections to services including S&P Global and FactSet. Its wider connector ecosystem includes more than 50 integrations, including Datasite, Box, Preqin and Intapp.
The product brings financial research, modeling and client-material creation into the same ChatGPT environment. OpenAI says teams can research across multiple sources, follow figures across reporting periods, interpret annotations in public financial data and turn the resulting analysis into spreadsheets, documents, slides, interactive charts and other artifacts.
“We’re effectively teaching ChatGPT to research like an analyst and back up its conclusions like an analyst as well,” OpenAI’s Head of ChatGPT Nick Turley said during a press roundtable accompanying the launch.
ChatGPT for Financial Services is available now to eligible financial institutions, according to OpenAI, with customers directed to contact the company or their existing account teams. The company has not publicly disclosed pricing, minimum seat requirements, geographic restrictions or the criteria it uses to determine which institutions are eligible.
For enterprise technology leaders, the significance is less that another AI model can summarize an earnings transcript or generate a pitch deck. Financial institutions increasingly have multiple models capable of doing those things.
OpenAI is instead trying to package the model, professional data, entitlements, firm templates, connectors and administrative controls into one governed environment where an analyst can move from research to analysis to a client-ready work product.
A goldmine of realtime financial data sources
ChatGPT for Financial Services includes premium datasets from providers including Daloopa, PitchBook and LSEG News, covering information such as earnings transcripts, financial statements, company fundamentals and private-company data.
The underlying idea of bringing professional financial data into ChatGPT is not new. PitchBook announced an OpenAI MCP connector in November 2025, LSEG announced its connector in December, and OpenAI made financial-data integrations a centerpiece of its March 2026 launch of GPT-5.4 and ChatGPT for Excel.
What changes with Thursday’s product is the packaging and architecture: selected professional datasets are now included in a dedicated financial-services product without separate data contracts or connector setup, according to OpenAI, and are indexed and hosted on OpenAI infrastructure. Those existing data relationships are also now being paired with GPT-6 Astra inside the finance-specific product.
Unlike a conventional connector that retrieves information from a third-party system at query time, OpenAI says it indexes and hosts this bundled data on its own infrastructure. Customers can begin using those datasets without negotiating separate contracts or configuring individual connectors, according to the company.
OpenAI argues that hosting and indexing the information itself allows it to improve retrieval accuracy, latency and citation behavior.
There are effectively three data-access paths inside the product. Selected premium datasets — including data from Daloopa, PitchBook and LSEG News — are bundled, indexed and hosted by OpenAI and require no separate contract or connector setup.
Separately, OpenAI is working with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody’s on shared sign-in integrations that let providers recognize a user through ChatGPT and apply the data entitlements that user already holds. A third category consists of MCP and other connectors, including optimized connections such as S&P Global and FactSet and the broader ecosystem of more than 50 connectors.
That distinction matters because a provider such as LSEG can appear in more than one category. The included LSEG News offering should not be interpreted to mean every LSEG dataset is bundled with ChatGPT for Financial Services; OpenAI separately describes LSEG as part of its entitlement strategy for licensed content.
That architectural decision could be important for firms evaluating the product. Generative AI in finance increasingly depends less on whether a model knows general financial concepts and more on whether it can reliably retrieve the exact table, footnote or disclosure that supports an analysis.
“It’s one thing to nominally connect to a bunch of stuff, and it’s another thing to really teach these models to be literate in what that data means and be able to retrieve over it,” Turley said.
For example, OpenAI describes a banker conducting a P&L normalization analysis who needs to inspect the reconciliation behind adjusted EBITDA, determine which costs management excluded and decide how those adjustments should affect a valuation. The objective is not simply to return an EBITDA number but to preserve enough provenance that a professional can inspect the evidence and make the judgment independently.
Data governance is key
OpenAI is pursuing a second model for data that institutions already purchase. The company says it is working with S&P Capital IQ, LSEG, MSCI, Factiva and Moody’s on shared sign-in and entitlement integrations. Under that system, data providers will be able to recognize a user through their ChatGPT identity and automatically expose information the person is already licensed to access.
That is potentially significant for enterprise deployments because financial-data licensing is rarely a simple organization-wide on/off switch. Individual desks, business units and employees can have different entitlements, and allowing ChatGPT to inherit those permissions could reduce both administrative friction and the risk that AI becomes an unintended path around existing access controls.
OpenAI is also optimizing MCP-based integrations with frequently used financial providers such as S&P Global and FactSet and says its broader connector ecosystem now exceeds 50 integrations, including Datasite, Box, Preqin and Intapp. The company says it has focused on reducing the tool-use and retrieval problems that can make generic MCP connections unreliable in production financial workflows.
The company is now putting numbers behind some of those reliability claims. In OpenAI’s evaluation, connector error rates fell from 5.09% to 1.99% for Quartr, from 6.84% to 2.66% for S&P Global, from 9.59% to 6.45% for FactSet and from 7.53% to 2.57% for Daloopa after its optimization work. OpenAI says those improvements came through automated evaluation and iteration.
The announcement does not disclose the size or composition of the test set or precisely define what counted as a connector error, so financial institutions would still need to validate reliability against their own workloads, permissions and queries.
Astra's intelligence is the driver
The model sitting above those data sources is GPT-6 Astra. OpenAI describes Astra as state of the art across three capabilities it considers central to financial-services work: information retrieval, financial reasoning and artifact generation.
The company says Astra can navigate figures, tables and supporting notes in financial documents, conduct financial analysis from those inputs and synthesize the work into documents, spreadsheets and presentations. ChatGPT for Financial Services will also receive newer OpenAI models as they are released rather than remaining tied permanently to Astra.
On OfficeQA Pro, an independently developed benchmark contains 133 questions over roughly 89,000 pages of U.S. Treasury Bulletins spanning nearly a century, GPT-6 Astra scored 69.9% correctness in OpenAI’s evaluation, compared with 60.2% for GPT-5.6 Sol and 62.4% for Claude Fable 5.1 — advantages of 9.7 and 7.5 percentage points, respectively.
OfficeQA Pro is particularly relevant to the product OpenAI is pitching because it was designed to test grounded enterprise reasoning rather than general knowledge.
But the result should still be read narrowly. The 69.9%, 60.2% and 62.4% figures are results reported by OpenAI for its evaluation setup; the launch post does not specify enough about the agent harness, reasoning settings or other test-time configuration to turn the chart into a universal ranking of the models for financial work. The OfficeQA Pro researchers themselves found performance can change based on factors including model choice, document representation, retrieval strategy and test-time scaling.
During a press briefing, Turley said the company believes its latest model is roughly twice as efficient as the strongest alternative when measured by cost per task, and said OpenAI has focused its evaluations on corporate-data retrieval, financial reasoning and the fidelity of generated artifacts.
Turley argued that those practical measures matter more to the intended customer than conventional model leaderboards. “Academic benchmarks are not interesting to this industry,” he said. “It is about real-world utility, and that’s on cost and on utility.”
The distinction matters for buyers comparing models on total cost of ownership. Cost per token alone says relatively little about an investment-banking workflow if one model requires more retries, manual corrections or analyst review than another.
Conversely, a more expensive reasoning run may still be cheaper overall if it can accurately complete a multi-step model or deck with less human rework.
OpenAI’s cost-per-task framing is therefore relevant, but financial institutions will need considerably more evidence before treating the claim as a procurement metric.
Artifact creation is another major part of the product
Administrators can publish approved Excel, Word and PowerPoint templates through a dedicated administration page. Once a firm’s templates and style guidance are configured, OpenAI says employees can generate valuation models, research notes and pitchbooks using the institution’s preferred formats.
That is also where OpenAI says a key distinction between visually impressive AI output and production-grade financial work emerges. “It’s very easy to make slides that look good, but it’s much harder to make the slides actually make sense, because the AI has to reason under the hood and see the right data,” Turley said.
During the accompanying demonstration, OpenAI also showed administrators centrally distributing templates while individual users could save templates for their own future use. In another demonstration, the system took financial-model outputs and created an interactive website with controls that allowed users to change assumptions and see how those changes flowed through projected results.
OpenAI technical staffer Joseph Kim, who demonstrated the product in a press briefing with VentureBeat, said OpenAI sees that kind of interface generation as more than a faster version of an existing workflow. “We’re not just trying to accelerate the current thing, but we’ve also been spending a lot of time thinking about new ways to leverage this platform to figure out new ways of doing things.”
That points toward a broader shift in the role of generative AI inside financial institutions. Instead of merely generating text inside a chat window, frontier systems are increasingly being asked to produce operational artifacts — spreadsheets, presentations, applications and interfaces — that employees continue editing in the software they already use.
OpenAI already laid groundwork for that strategy earlier this year with ChatGPT for Excel and financial-data integrations. Its March release brought financial modeling directly into Excel and added integrations for financial datasets and enterprise sources. OpenAI said at the time that ChatGPT could build and update models, reason across workbooks and trace changes back to cells and formulas.
OpenAI is far from the only AI firm targeting finance
Anthropic launched Claude for Financial Services in 2025, combining Claude with financial-data connections and enterprise deployment support, and this year expanded the offering with finance-specific agents for workflows including pitchbook generation and KYC screening. It has also pushed Claude across Excel, PowerPoint, Word and other Microsoft 365 applications.
Microsoft, meanwhile, is building many of the same financial-data sources into Copilot. Its financial-services strategy includes connectors from providers such as LSEG, Moody’s, Daloopa, FactSet, PitchBook and S&P Global while keeping AI inside Excel, Teams, Outlook and other Microsoft 365 applications where financial employees already work.
Incumbent financial-data providers are moving in the same direction. FactSet introduced FactSet AI for Banking with Finster AI in March as a secure environment for automating investment-banking and research workflows, while S&P Global Market Intelligence announced a partnership with Farsight in July to generate pitch decks, confidential information memoranda and valuation materials from within Capital IQ Pro.
That competitive landscape changes the buying question. Financial institutions no longer have to decide whether they want AI-assisted financial research. Increasingly, they must decide where that intelligence should sit: inside the data platform, inside Microsoft Office, inside a standalone frontier-model platform, or across a combination of all three.
OpenAI’s answer is to make ChatGPT Work the orchestration layer while allowing professional data and enterprise systems to flow into it.
Turley made that ambition unusually explicit during the briefing: “When it comes to end-user productivity, when it comes to the way that analysis happens on the ground, this is the canonical product that we hope the industry adopts.”
That statement helps frame ChatGPT for Financial Services as more than a specialized SKU. OpenAI is effectively competing for the interface through which bankers conduct research, manipulate data, build models and produce client work — even as those artifacts ultimately continue into Excel, PowerPoint and other established enterprise systems.
Administrators choose the workspace allowances
Governance will be critical if that architecture is going to extend beyond pilots. ChatGPT for Financial Services builds on ChatGPT Enterprise controls including SAML SSO, SCIM provisioning and role-based access controls. Administrators can control access to skills and apps by role and determine which supported integrations have read or write permissions.
Organizations can create multiple workspaces to enforce information barriers. That could be particularly relevant for institutions that need to separate teams handling different clients, transactions or sensitive information.
OpenAI says business data is encrypted at rest and in transit, workspace retention can be configured by administrators, and supported workspace logs can be exported through the OpenAI Compliance Platform for use in existing audit and investigation processes. OpenAI separately states that data shared with ChatGPT Enterprise is not used to train or improve its models by default.
The company’s own executives repeatedly returned to deployment constraints during the briefing. Discussing requirements such as governance and administration, Turley put it starkly: “You can have superintelligence; it just doesn’t matter” if the institution cannot actually deploy it.
Those controls arrive as regulators pay closer attention to how banks give AI systems access to sensitive data and operational systems. U.S. banking regulators have been examining issues including AI data access, governance, third-party risk and human oversight as adoption spreads into more consequential workflows.
Banks are already moving quickly. Institutions including Morgan Stanley, BNY, UBS, Goldman Sachs, JPMorgan and Citi have been deploying or testing agentic systems across wealth management, trading, treasury and other functions, while executives continue to emphasize human oversight for higher-risk activity.
For enterprise decisionmakers, ChatGPT for Financial Services therefore looks less like the beginning of AI adoption on Wall Street than the beginning of a platform contest over where that adoption consolidates.
OpenAI now has a more finance-specific answer to several questions that previously complicated large deployments: how professional data enters the model, how existing entitlements carry over, how firm templates can be centrally distributed, how separate business populations can be isolated and how model-generated research becomes a usable Excel model, document or pitchbook.
But several procurement-level questions remain
OpenAI has not disclosed pricing or whether ChatGPT for Financial Services carries a premium over existing enterprise agreements. It has not publicly specified minimum deployment sizes, detailed eligibility criteria or geographic availability. It has not explained the economics of the premium datasets included with the service, nor how those economics may affect pricing as coverage expands. And it has not published enough detail to independently validate its claims around Astra’s comparative financial performance or cost per task.
Institutions will also need to determine how OpenAI’s hosted-data approach fits their own data-governance policies, how information barriers behave across connectors and generated artifacts, what controls apply when agents receive write permissions, and how organizations should validate model or data-provider changes over time as newer frontier models are automatically introduced into the product.
Those questions do not negate the significance of the launch. They define the next phase of evaluation.
The strategic proposition is increasingly clear: OpenAI does not simply want bankers to ask ChatGPT questions. It wants ChatGPT to become a governed workspace where professional financial data is retrieved, analyzed, modeled and converted into the work products that move through an institution.
Whether banks consolidate around that environment will depend less on the quality of a demo than on provenance, permissioning, reliability, interoperability and the measurable cost of completing real work.
We have reached out to OpenAI for additional information on pricing, eligibility and rollout, premium-data licensing, model evaluation methodology and enterprise governance details, and will update this story when we hear back.