OpenAI launched ChatGPT for Financial Services (announcement) on September 10–11, 2026 — a tailored ChatGPT Work experience for investment banking and equity research, powered by GPT-6 Astra, with premium market data hosted inside OpenAI's stack and firm-owned Excel/Word/PowerPoint templates. Same week as the ChatGPT Work Data agent; different product. That one wires your warehouse and BI tools. This one is a finance vertical SKU: bundled datasets, citation tracing, pitchbook/model artifacts, and sales-gated enterprise access.
TL;DR
| Question | Answer |
|---|---|
| What launched? | ChatGPT for Financial Services — ChatGPT Work SKU + GPT-6 Astra + finance data/templates |
| Who can buy it? | Eligible FIs via OpenAI sales / account team; needs ChatGPT enterprise — no public seat price |
| Design partners | Morgan Stanley, Evercore |
| Bundled data (OpenAI-hosted) | Daloopa, PitchBook, LSEG News, Crunchbase |
| Connect your own | Shared entitlements (S&P Capital IQ, LSEG, MSCI, Factiva, Moody's) + 50+ MCP connectors (FactSet, S&P Global, Datasite, Box, Preqin, Intapp, …) |
| Artifacts | Editable models, research notes, pitchbooks using firm Excel/Word/PPT templates |
| Citations | Trace figures/claims to specific paragraphs/tables; preview supporting passage |
| vs Data agent | Market-data + IB/ER workflows, not Snowflake/BI warehouse chat |
Who can get it — and what do you pay?
This is not a Plus/Pro toggle and not a self-serve public SKU. OpenAI says eligible financial institutions get it through sales or their account team. Press coverage and OpenAI's own framing agree on the gate: you need a ChatGPT enterprise account, and commercial terms are custom. No public per-seat price was announced.
For builders and IT buyers, that means: budget via enterprise renewal / expansion with OpenAI, not a line item you can copy from a pricing page. If you already run ChatGPT Work or Enterprise, ask your account team whether your org is in scope and what the finance SKU adds to the existing Work bill — do not assume Data agent access or a warehouse connector equals this vertical.
What's included vs ChatGPT Work and the Data agent?
| ChatGPT Work (base) | Data agent (Sept 10) | Financial Services SKU | |
|---|---|---|---|
| Audience | General enterprise knowledge work | Ops/sales/analysts querying internal data | IB / equity research (launch scope) |
| Model emphasis | Work + Codex tool surface | Same Work surface + data plugin | GPT-6 Astra for financial retrieval/reasoning |
| Primary data | Your Drive/Slack/etc. connectors | Snowflake, BigQuery, BI tools | OpenAI-hosted market data + firm subscriptions |
| Signature output | Docs, code, browser tasks | Dashboards / approved actions | Models, tearsheets, pitchbooks in firm templates |
| Access | Business/Enterprise Work | Plugin on Work | Sales-gated vertical on Enterprise |
The Data agent post covers warehouse permissions and prompt-injection risk when anyone can ask SQL in English. Do not conflate the two. Financial Services is about market data plumbing, citation-backed research, and client-ready artifacts — not replacing Tableau against your warehouse.
ChatGPT Work's broader tool/skill surface (documented in Simon Willison's reference dump and the Work vs Codex guide) still matters: this SKU sits on that harness, with finance-specific data and admin template publishing on top.
How do citations and "trace the figure" actually work?
OpenAI's pitch is audit-friendly research: when the model cites a number, you should be able to jump to the specific paragraph or table and preview the supporting passage — not a vague "according to filings" footnote.
That is the right product instinct for regulated workflows. It is not a substitute for model-risk and compliance review. Treat citation UI as:
- Faster source check — open the table/passage behind adjusted EBITDA or a comps row.
- Still human-gated for client work — pitchbooks and research notes that leave the firm need the same review you would apply to a junior analyst's draft.
- Dependent on OpenAI's index — quality tracks how cleanly Daloopa/PitchBook/LSEG/Crunchbase (and connected entitlements) are retrieved, not magic immunity to hallucinated math between cited cells.
OpenAI also cites OfficeQA Pro (finding/analyzing information across U.S. Treasury Bulletins with tables, charts, footnotes): GPT-6 Astra 69.9% vs GPT-5.6 Sol 60.2% — OpenAI's own number, no independent re-run in the announcement.
Templates: Excel, Word, PowerPoint in the firm's format
Admins publish firm Excel, Word, and PowerPoint templates so valuation models, research notes, and pitchbooks land in house style. Demo coverage (e.g. Nick Turley walking an M&A target into a formatted deck) matches what deal teams actually ship: comps, sensitivity, and slides that look like the bank's template, not a generic AI export.
Practical implication for builders: template governance becomes part of the AI rollout — who may publish templates, which workspaces get which decks, and how you version them when branding or disclosure language changes.
Astra vs Sol for this SKU
The Financial Services product is powered by GPT-6 Astra, which OpenAI positions for retrieval across financial tools, financial reasoning, and accuracy of generated content. On the Astra launch and Astra vs Fable 5.1 coverage: Astra sits at the top of OpenAI's Astra > Sol > Terra > Luna scale; Sol remains the prior GPT-5.6 flagship tier.
For this vertical, the useful comparison is not "is Astra the Intelligence Index leader?" — it is whether Astra's retrieval + artifact loop is good enough under your compliance bar with your templates and your connected entitlements. Use independent benchmarks for API/model choice; use a pilot deal team for the SKU.
MCP connectors and existing Bloomberg/FactSet-style subscriptions
Beyond the four bundled datasets OpenAI indexes and hosts, the announcement describes:
- Shared sign-in / entitlement work with providers such as S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva, and Moody's so existing subscriptions recognize the ChatGPT login.
- ~50+ MCP connectors, with optimized paths called out for heavy-use tools like FactSet and S&P Global, plus ecosystem names such as Datasite, Box, Preqin, and Intapp.
If you are wiring custom finance tools, the same MCP fundamentals apply as elsewhere on explainx.ai: least-privilege connectors, clear auth boundaries, and no assumption that "connected" means "safe to act without review." Work's cloud browser for signed-in sites is a separate surface — useful for some research flows, but not a replacement for entitlement-aware data connectors.
Press summaries sometimes name Bloomberg alongside FactSet-class terminals; prefer OpenAI's listed partners and MCP set when scoping a procurement checklist, and confirm entitlement behavior with your account team.
Compliance caveats — verify outputs
Enterprise controls ride on ChatGPT Enterprise: SAML SSO, SCIM, RBAC, encryption at rest/in transit, configurable retention, and Compliance Platform log export. Multiple workspaces can support information barriers (MNPI-aware segregation). Business data is not used to train OpenAI models by default, per OpenAI's enterprise terms as described in the launch materials.
None of that absolves:
- Model error between cited facts — citations reduce invented sources; they do not guarantee correct arithmetic or correct comps selection.
- Over-trust on client materials — treat generated pitchbooks like analyst drafts.
- Connector sprawl — each MCP or entitlement path widens what a chat can pull; scope roles the same way you would for the Data agent's service accounts.
What this means for what you build or pay
| If you… | Practical consequence |
|---|---|
| Build internal IB/ER copilots | Evaluate whether OpenAI-hosted Daloopa/PitchBook/LSEG/Crunchbase + firm templates replace a pile of custom RAG + MCP glue — or only cover a subset of your data stack. |
| Already pay ChatGPT Enterprise / Work | This is an add-on vertical, sales-quoted — not free with Work, and not the same SKU as the Data agent plugin. |
| Already pay FactSet / S&P / LSEG / etc. | Plan entitlement SSO and MCP enablement with legal/compliance; bundled OpenAI-hosted data does not automatically replace every terminal subscription. |
| Ship agent products on the API | OpenAI points specialized apps at the API; the Work SKU is the end-user research/pitch surface. Separate build vs buy. |
| Own model-risk / compliance | Budget for citation review workflows, workspace barriers, template admin, and log export into existing audit tools — not just seat licenses. |
Launch scope is investment banking and equity research (comps, valuation sensitivity, pitchbooks), shaped with Morgan Stanley and Evercore. OpenAI says partner feedback will drive post-training and later expansion into other FSI categories — so treat "financial services" branding as IB/ER first, not a claim that every banking workflow is covered on day one.
Related on explainx.ai
- OpenAI Data agent in ChatGPT Work — enterprise safety — sibling Sept 10 launch (warehouse/BI, not market data)
- GPT-6 Astra launch: benchmarks, pricing, rollout
- GPT-6 Astra vs Claude Fable 5.1
- ChatGPT Work vs Codex — complete guide
- Simon Willison's ChatGPT Work tools & skills reference
- ChatGPT Work cloud browser — signed-in websites
- What is MCP? Model Context Protocol guide
Primary source: OpenAI — Introducing ChatGPT for Financial Services
Feature lists, partner names, OfficeQA Pro scores, and access rules reflect OpenAI's September 10–11, 2026 announcement and contemporaneous press summaries. Commercial packaging and connector entitlements change — confirm current availability and terms with OpenAI before procurement or production rollout. Follow @explainx_ai for ChatGPT Work coverage.