On 2 June 2026, OpenAI unveiled six role-specific plugins for its agent Codex, two of them close to capital markets: Investment Banking and Public Equity Investing. Together the plugins connect the agent to 62 business applications and start with 110 prebuilt skills. For the Public Equity Investing plugin, OpenAI names prominent data partners: Moody's, FactSet, S&P, the London Stock Exchange Group (LSEG), PitchBook and Hebbia, supplemented by Daloopa and Datasite. It is OpenAI's most visible push into the financial sector, and it meets a competitor already in the market.
Because Anthropic had led. As early as January 2026, the company launched eleven role-specific plugins for its working environment Claude Cowork; in May a dedicated suite of around ten prebuilt finance agents followed, together with the then flagship model Claude Opus 4.7. It would be tempting to read OpenAI's June announcement as a direct response. It is only partly that: OpenAI itself does not name Anthropic as the trigger, and the immediate forerunner of the plugin logic was rather the January launch of Claude Cowork. More precisely, the sober observation is that the Codex push joins an already running race rather than starting it.
What: OpenAI unveils role-specific Codex plugins, including Investment Banking and Public Equity Investing
When: 2 June 2026; Anthropic's finance agent suite already on 5 May 2026
Scope: 62 connected business applications, 110 prebuilt skills across all six plugins
Usage: According to OpenAI, more than 5 million weekly Codex users, around 20 per cent without a developer role (self-reported, not externally audited)
Relevance: Model strategy, data connectivity and governance of agentic systems in the front and middle office
What OpenAI unveiled on 2 June
The Codex plugins turn a coding agent into a role-specific work assistant. For investment banking that means automating recurring analytical and preparatory work; for public equity investing, connecting to financial databases. One important qualification is often lost in the excitement: the 62 applications and 110 skills are a total across all six plugins, not the equipment of the investment banking plugin alone. And the prominent list of data partners is explicitly assigned to the Public Equity Investing plugin; for the investment banking plugin, OpenAI stays with the vaguer wording of trusted data.
OpenAI flanks the push with usage figures: more than five million weekly Codex users, up from around 600,000 at the start of the year, roughly 20 per cent of them without a classic developer role and growing three times as fast. These figures come solely from OpenAI's own account; independent telemetry is not available. Anyone folding them into an investment decision should label them as a self-reported figure, not an audited metric.
The race has two vendors and one data tempo
What distinguishes the two vendors is less capability than cadence. Anthropic equipped its finance agents early with deep data integrations: Moody's as a native application in the model context, full connection to Microsoft 365, plus a string of further data connectors. Named clients range from JPMorgan, Goldman Sachs and Citigroup through Visa to Citadel, the Bank of New York and FIS. OpenAI counters with reach and a horizontal role strategy across six professions.
The decisive point for banks runs across this competition: the same data providers appear at both vendors. Moody's and FactSet connect to both model ecosystems rather than binding themselves exclusively. For a bank's vendor strategy that is good news. Data connectivity is not chained to a single language model. Anyone drafting their data contracts cleverly can run agents from OpenAI and Anthropic in parallel or in rotation without an exclusive data access becoming a barrier.
The real risk: the model rotates faster than validation
The underestimated operational finding sits in a detail of the timeline. Anthropic presented Claude Opus 4.7 on 5 May, explicitly as leading on financial tasks. Only three weeks later, on 28 May, Claude Opus 4.8 replaced the model, according to the maker around four times less likely to let its own coding errors pass unflagged. Three weeks lay between two flagship models.
For model risk management this is a structural break. Classic validation cycles, as supervisory practice around the principle of sound model governance expects, are designed for months, not weeks. When the base model behind an agent changes faster than internal validation can keep up, a gap opens between the validated and the productive state. Banks therefore need their own control point for the model version, independent of the vendor, that registers every change and triggers a renewed, proportionate review.
Governance: factual accuracy and auditability in the front office
Both suites target documents with supervisory relevance: pitchbooks, comparable analyses, credit memos. It is precisely there that the two classic weaknesses of agentic systems are most dangerous. The first is factual accuracy: an agent pulling financial metrics from several sources can combine numbers plausibly but wrongly. Neither OpenAI nor Anthropic publishes robust error rates for the extraction of financial data. The second is auditability: when an agent orchestrates 62 applications, the seamless traceability of which data source shaped which output becomes a challenge. Both are not marginal questions but the core of every model approval in a regulated environment.
EU AI Act: classify the role, not the vendor
When it comes to regulatory classification, differentiation is decisive. The European regulation on artificial intelligence, the EU AI Act, classifies creditworthiness assessment and insurance risk assessment as high-risk under Annex III. Most front-office applications in capital markets, such as producing a pitchbook or a comparable analysis, do not automatically fall under it. Borderline cases are agents close to anti-money-laundering and customer due diligence, such as Anthropic's customer-screening agent or the Claude-based financial crime agent from FIS. The European Commission presented draft guidelines on high-risk classification on 19 May 2026 and consulted until 23 June; the so-called Digital Omnibus also shifts the central Annex III deadline from 2 August 2026 to 2 December 2027. For banks it follows that risk classification should be carried out per agent role, not blanket per vendor.
Recommendations for operational practice
For institutions introducing agents in the front and middle office, the vendor race is less a purchasing decision than a steering task. Five fields of action stand out.
Short term: Since the same data providers connect to both model ecosystems, contracts with Moody's, FactSet and others should be negotiated so that parallel or rotating operation of OpenAI and Anthropic agents does not fail on exclusive access clauses. This decoupling is negotiating leverage.
Immediate: The three-week cadence from Claude Opus 4.7 to 4.8 shows that models change faster than classic validation cycles. Governance frameworks need a vendor-independent control point that registers every model change and triggers a proportionate revalidation.
Before go-live: A pitchbook agent is, in regulatory terms, something different from a customer-screening agent. Classification under the EU AI Act belongs to the individual role, not the vendor, precisely because the Commission guidelines are not yet final.
Ongoing: User figures and benchmark claims come predominantly from the vendors' own accounts. Anyone folding them into business cases should label them as such and, where possible, test them against their own pilot measurements.
Before go-live: Agents preparing financial metrics for pitchbooks or credit memos need controls for source evidence, number checking and seamless traceability. Without these controls, the efficiency gain becomes a reputational and supervisory risk.
Christian Schablitzki
Strategy & Management Consultant · Agentic AI expert for financial institutions
More than 20 years in investment banking and derivatives trading, followed by over 10 years advising financial institutions. Currently a Partner at Infosys Consulting in Germany. Certified in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.
LinkedIn profile →Keep reading – every 14 days in your inbox.
Capital-markets insights, regulatory updates and AI trends. Concise, well-founded, free of charge.
GDPR-compliant. Unsubscribe at any time.