The model gets cheap, the switch does not: what ChatGPT for Financial Services is actually selling
If capability keeps getting cheaper, a lab cannot defend its margin on the model. It can defend the place the model occupies inside your company. Here is what OpenAI announced on 10 September, and the two verified numbers that cut both ways.
OpenAI launched ChatGPT for Financial Services on 10 September 2026, built with Morgan Stanley and Evercore, aimed at investment banking and equity research. It ships with built-in financial datasets, citations traceable to a source line, templates an administrator publishes for the whole firm, single sign-on and role-based access. The interesting part is not the model inside it. It is what the product is trying to make expensive.
The premise behind that strategy is measurable, not rhetorical. Epoch AI tracks the price of reaching a fixed capability level and finds it falling between 9 and 900 times a year depending on the benchmark: general knowledge at GPT-3 level cost $60 per million tokens in November 2021 and $0.18 in February 2025. If what you sell is the model, your price has a short half-life.
So you sell what surrounds it instead: connected data, house formats, access rights, validated procedures, trained teams. A rival can match the capability and still have replaced none of that. This is the oldest moat in enterprise software. What makes this case worth checking is that the spending data shows enterprises switching models anyway, and that the data OpenAI wired in is sold to its competitor too.
What was announced, what the numbers verify, and where the moat is thinner than it looks
1. What OpenAI actually shipped on 10 September 2026
A ChatGPT Enterprise tier for financial institutions, shaped by a design partnership with Morgan Stanley and Evercore, with investment banking and equity research named as the starting point. Reported features: GPT-6 Astra as the reasoning model, built-in datasets covering earnings transcripts, financial statements, company fundamentals and private companies, granular citations that trace a figure back to its source, Excel, Word and PowerPoint templates an administrator publishes firm-wide, SAML single sign-on, SCIM provisioning, role-based access controls, workspace encryption and compliance log export. It is sold only to eligible institutions that contact OpenAI, and no price has been published.
1. What OpenAI actually shipped on 10 September 2026: verified pricing, fit and cautions →
2. The falling-price premise is right, and it is measured
Epoch AI tracks the cheapest way to reach a fixed benchmark score over time, which is the honest way to ask what intelligence costs. Its range is 9x to 900x per year depending on the milestone. On general knowledge, GPT-3 cost $60 per million tokens in November 2021, GPT-3.5 Turbo $2.00 in March 2023, Gemini 2.0 Flash $0.18 in February 2025. On coding, GPT-4-0314 cost $37.50 per million in March 2023 and an 8-billion-parameter Llama 3.1 matched that score at $0.10 in July 2024. Stanford's AI Index measured the same effect on a different cut: $20 per million for GPT-3.5-level output in November 2022, $0.07 by October 2024. Epoch itself cautions that the steepest drops are the most recent ones and may not persist.
3. The gap today, from our own verified records
Our knowledge base has OpenAI's flagship API tier at $5.00 input and $30.00 output per million tokens, checked 27 July 2026. DeepSeek V4.1 Flash, open weights under MIT, publishes $0.15 to $0.30 input and $0.60 to $1.20 output. That is a 20x to 30x spread on list price. The honest qualifier is that it is not yet the same purchase: on the hardest current benchmarks our review of that model found it losing to Claude Opus 5.0 by 12 points on Terminal-Bench 4.0 and by 20 on Humanity's Last Exam. The cheap option is close, not equivalent.
3. The gap today, from our own verified records: verified pricing, fit and cautions →
4. The number that says enterprises do switch
Menlo Ventures, which surveys enterprise AI budgets annually, put OpenAI's share of enterprise LLM API spend at roughly 50% in 2023 and 27% in 2025, while Anthropic went from 12% to 24% to 40% over the same period. Whatever lock-in exists at the raw API layer, it did not stop a market leader from losing about half its share in two years. That is the pressure a workflow product is built to answer, and it is the strongest evidence that the model by itself is not a defensible position.
5. And the number that says they mostly do not
The same Menlo report found the opposite one layer up: once an enterprise has chosen a vendor, it tends to stay, and even where switching costs are low, most teams simply upgrade to their existing provider's newest model rather than evaluate a rival. Both findings are true because they describe different buyers. A platform team swapping an API endpoint behind a gateway moves easily. A research desk with published templates, entitlement mappings and a validated review procedure does not.
6. What is actually being sold here is not intelligence
Read the announced feature list as a migration bill and it becomes clear. Firm-published document templates have to be rebuilt. Role-based access and SCIM provisioning have to be re-mapped to the same people. Compliance log export has to satisfy the same auditor. Shared entitlements let the product read the S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's subscriptions the bank already pays for, so the plumbing has to be reconnected seat by seat. None of that is capability. All of it is work a competitor's better model does not do for you.
7. The data layer is not the moat, because it is not exclusive
Anthropic launched Claude for Financial Services on 15 July 2025, fourteen months before OpenAI's product, with Daloopa, PitchBook, FactSet, S&P Global, Morningstar, Databricks, Snowflake, Box and Palantir as named connectors. Two of the three datasets OpenAI built in, Daloopa and PitchBook, were already on that list. Anthropic followed on 5 May 2026 with pre-built agents, Microsoft 365 integration and Moody's data, reported alongside a JPMorgan partnership, and on 14 September 2026 with a product for financial advisers connecting to BlackRock, Charles Schwab and Addepar. The vendors sell to both sides. What differs is the workflow built on top.
8. Why the commercial pressure is real, in OpenAI's own numbers
OpenAI's CFO Sarah Friar told investors on 14 August 2026 that enterprise revenue had overtaken consumer, from a 60-40 split favouring consumer at the start of the year, and said the lines crossed faster than expected. Sacra estimates the company reached about $40 billion of annualised revenue in July 2026, roughly double its end-2025 figure, though that is an outside estimate and not an audited disclosure. A consumer subscription churns in a month. An installed research workflow at a bank does not, which is exactly why the money is moving.
9. One fact a reader should simply have
Morgan Stanley is the design partner named on the product and, per reporting in June 2026, one of the two top bookrunners selected for OpenAI's confidential IPO filing alongside Goldman Sachs. No lead-left designation had been awarded at that point. We are not claiming the two facts are connected, and there is no evidence that they are. But when a launch cites a bank as co-designer, it is worth knowing that the bank also stands to earn fees on the vendor's listing.
The argument that falling compute costs could widen a lab's margin rather than cut its price holds only under outcome pricing, where the client pays for the analysis rather than the tokens. OpenAI has not published a price for this product, so that part cannot be verified, and the entry condition is a ChatGPT Enterprise contract, which is a seat model. Treat the margin story as a plausible direction, not an announced one.
And the question of whether the labs eventually bypass their clients has a narrower answer than it sounds. This product's stated starting point is research, valuation models and client materials, the work a junior banker does. The substitution is not beginning at the top of the bank, it is beginning inside it, and the firms buying the product are the ones choosing who it substitutes for.
Where we track this
- One AI, two price tags: rent versus irrigation →
The macro version of this argument. Why the American stack has to charge for intelligence while the Chinese stack works to make it nearly free, with the capex numbers behind both.
- The cheapest serious model, and its full benchmark table →
The concrete $0.15 alternative referenced above, including the benchmarks where it still loses to the frontier by twenty points.
- ChatGPT in our tool library →
The knowledge-base record: verified pricing, the data-handling default and the sources behind both.
Questions people ask
- What is ChatGPT for Financial Services?
- A version of ChatGPT Enterprise for financial institutions, launched 10 September 2026 and designed with Morgan Stanley and Evercore. It combines GPT-6 Astra with built-in financial datasets from Daloopa, PitchBook and LSEG News, source-level citations, firm-published Excel, Word and PowerPoint templates, single sign-on, role-based access and compliance log export, plus connectors to data subscriptions a firm already holds. It is sold only to eligible institutions and the price is not published.
- Does using it lock my firm in?
- Not contractually, as far as anything published shows, but practically it raises the cost of leaving. The templates, access mappings, entitlement connections and validated procedures are yours to rebuild if you move. That is ordinary enterprise software economics: a rival can match the model and still leave you with weeks of migration. The defence is to keep the expensive parts portable, meaning your prompts, your evaluation set and your data connections documented independently of the vendor.
- Is OpenAI's financial data exclusive to OpenAI?
- No. Daloopa and PitchBook, two of the three built-in datasets, were already named connectors in Anthropic's Claude for Financial Services, launched 15 July 2025. The broader integrations rely on subscriptions the firm already pays for, such as S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's. The data providers sell to every lab. The differentiation is in the workflow, the citations and the controls built on top of them.
- If inference keeps getting cheaper, will my AI bill fall?
- Only if what you buy is tokens. Epoch AI's measured 9x to 900x annual price decline applies to reaching a fixed capability level, and in practice teams spend the saving on harder tasks, longer reasoning and more calls rather than banking it. A per-seat or per-outcome enterprise product breaks the link entirely: the vendor's cost per token can fall without your invoice following it down.
Written 19 September 2026 from an owner-supplied French commentary on OpenAI's financial-services launch; every claim was checked against reporting and primary documentation on 19 September 2026. OpenAI's own announcement page returned HTTP 403 to our fetches, so the product's feature list is reconstructed from two independent reports that quote it (CNBC, 10 September 2026, and a close paraphrase published by Unite.AI) and is labelled reported rather than confirmed throughout. Those two accounts differ on one point: the built-in datasets are given as Daloopa, PitchBook and LSEG News, while a wider list including Bloomberg, FactSet and Crunchbase appears to describe connectors used with a firm's own entitlements; we have reported the distinction rather than merge the lists. Corrections and additions to the source material: it names PitchBook and Daloopa as the integrated data, omitting LSEG News, the third built-in set; it presents the move as opening a new front, whereas Anthropic shipped Claude for Financial Services on 15 July 2025 with Daloopa and PitchBook already among its connectors, added pre-built agents and Moody's data on 5 May 2026, and launched an advisers product on 14 September 2026; and its illustrative rival offering the same work at a fiftieth of the price with equivalent reliability does not yet exist, since the cheapest serious model we have reviewed still trails the frontier by 12 to 20 points on the hardest benchmarks. Price-decline figures are Epoch AI's inference price trend data and Stanford's AI Index; the per-token comparison uses our own knowledge-base records checked 27 July 2026 and DeepSeek's published table. Enterprise API share is Menlo Ventures' 2025 enterprise survey, a sampled estimate rather than audited revenue. The revenue crossover is Sarah Friar's 14 August 2026 remarks as reported by CNBC; the $40 billion annualised figure is Sacra's estimate, not an OpenAI disclosure. Morgan Stanley's bookrunner role is June 2026 reporting on a confidential filing and remains subject to change. The margin argument in the closing is analysis, not an OpenAI statement: no price for this product has been published. Disclosure: TaskNorth's knowledge base recommends Claude models, and Anthropic, a direct competitor in this market, makes the models used in building this site.
Trying to work out which AI tools fit your task? Describe the outcome and get a Blueprint: the tools, the prompt, and the steps, with pricing we verified ourselves.