OpenAI Puts ChatGPT on Wall Street With 50 MCP Connectors

OpenAI has shipped ChatGPT for financial services, built with input from Morgan Stanley and Evercore and wired to roughly 50 MCP connectors.

The product runs on OpenAI's GPT-6 Astra line and is available to eligible institutions on ChatGPT Enterprise, according to a Dev.to breakdown by iFynx Studio citing Fortune's coverage of the launch. The interesting part for engineering teams is not the model tier — it is the connector list and what OpenAI chose to put around it.

What actually shipped

Per the same report, the launch bundles "dataset connectors (Bloomberg, FactSet, and preloads such as Daloopa, PitchBook, Crunchbase, LSEG News), roughly 50 MCP connectors, detailed citations, and generation of PowerPoint, Excel, and simple dashboards—with enterprise controls for SSO, SCIM, RBAC, and retention."

Read that list in order. Two of the four groups are data plumbing, one is provenance, and one is output format. The model is assumed.

MCP is doing the heavy lifting

Fifty connectors is the number to sit with. A protocol most teams first met through developer tooling is here carrying the integration load for a regulated product that has to name its sources. The source argues the pattern directly — "MCP is becoming the integration fabric. Prefer vendors and internal tools that speak MCP with capability cards" — and that is the author's recommendation, not an OpenAI commitment.

If you maintain internal tools that an assistant will eventually call, note what the interface is here: MCP, not a bespoke plugin API. That is a procurement signal worth raising at the next vendor review.

Which GPT-6 tier is under it

Astra sits at the top of the GPT-6 line. A separate Dev.to analysis of this week's releases notes the timing: "Opus 5.5 and GPT-6 dropped on the same day, September 22," with "GPT-6 Sol and Luna alongside GPT-6 Astra at the top tier," and characterises the field as "GPT-6 Sol is clearly a rung below GPT-6 Astra."

That matters for anyone pricing a comparable build. The finserv product is on the most expensive tier available, and the iFynx write-up flags a latency consequence — it recommends draft-versus-approved states for generated artifacts "especially when Astra-class generation takes minutes, not milliseconds." Treat that as the author's field observation rather than a published benchmark.

What to copy if you are not an eligible institution

Most teams reading this will not clear OpenAI's eligibility gate. The transferable parts are cheap to build and do not need a frontier model:

• Citation-first answers for any figure that could end up in a decision document — the source's blunt framing is that "research agents without citation UX and permissioned connectors are compliance liabilities."
• Connector admin, so security approves which sources an agent can read instead of discovering them later.
• Draft versus approved states on anything the agent generates for an external reader.
• A capability card per tool — the source's checklist is "purpose, inputs, outputs, risk tier, human confirmation rule, kill switch, and owner."

Our read: the eligibility gate is the least copyable thing here and the citation UX is the most. A permissioned connector catalogue and a source-visible answer format are ordinary engineering work, and they are what makes an agent's output auditable regardless of which model sits behind it.

What to watch

Three things will tell you whether this becomes a template or stays a Wall Street pilot: whether the connector catalogue opens beyond the named finance vendors, whether eligibility widens past the initial institutions, and whether the MCP connector count grows in the next release. If connectors keep accumulating faster than model tiers, the integration layer — not the model — is where the next round of enterprise AI competition gets decided.