Wall Street has been told for three years that AI will take over the junior work. On **10 September 2026** OpenAI shipped the product version of that claim - and, unusually, shipped a number with it.
The number is 69.9%. It is the best public score any model has on the financial document test OpenAI chose to publish. It also means roughly three answers in ten are wrong. Both statements are true at the same time, and everything useful about this launch sits in the gap between them.
What OpenAI shipped on 10 September
ChatGPT for Financial Services is a tailored version of ChatGPT Work built on GPT-6 Astra. It was shaped with **Morgan Stanley** and **Evercore** as design partners, who named two pain points: getting reliable access to data, and producing the artifacts - models, decks, memos - that the work actually ends in.
The product answers both directly.
Premium data is bundled in, indexed and hosted by OpenAI rather than reached through a connector: **Daloopa**, **PitchBook** and **LSEG News**. On top of that sit sign-in entitlements for **S&P Capital IQ**, **LSEG**, **MSCI**, **Factiva** and **Moody's**, optimised connectors for **S&P Global** and **FactSet**, and more than fifty further integrations including Datasite, Box, Preqin and Intapp.
The task list is explicitly the junior analyst's day: pull company and market research with a citation trail, reconcile adjustments, build financial models and spreadsheets, run P&L and EBITDA work, and turn the result into decks, memos and pitch books.
Pricing is not published. Neither are the eligibility rules. It is available to "eligible financial institutions" who contact OpenAI, which in practice means the buyer already has an account team.
The one number worth arguing about
OpenAI published a score on **OfficeQA Pro**, a benchmark of 133 questions asked over roughly 89,000 pages of US Treasury documents.
| Model | Correct | | --- | --- | | GPT-6 Astra | 69.9% | | Claude Fable 5.1 | 62.4% | | GPT-5.6 Sol | 60.2% |
OpenAI also states the system is around twice as efficient as alternatives when measured by cost per task. That is a claim, not an independent measurement, and there is no public method attached to it yet.
Take the benchmark result at face value and it is genuinely a good result. It is the highest published score on that set, and the gap to the next model is about seven points, which is not noise. Publishing a number at all puts OpenAI ahead of most enterprise AI launches, which ship adjectives.
But 69.9% is not a grade. It is a rate, and a rate has to be applied to the work.
What a 69.9% feels like at a desk
A single question answered correctly seven times in ten is a useful assistant. A deck is not a single question.
Suppose a pitch book carries twenty pulled figures and each one is an independent draw at 69.9%. The chance that all twenty are right is about 0.08%. Five figures, and it is about 17%.
Those numbers are an illustration, not a forecast - extraction errors are not independent, and a model that reads a filing correctly once tends to read it correctly again. The real distribution is much better than the arithmetic above. But the direction is the point: error rates compound across a document, and the output that reaches a client is a document.
There is a second effect that matters more. The errors are not scattered evenly across easy and hard questions. They concentrate where the document is ambiguous, the definition is non-standard, or the figure was restated - which is exactly the set of questions a junior analyst would have escalated to someone senior instead of answering.
The failure mode is not a model that gets basic things wrong. It is a model that is confident and fluent on precisely the questions that needed a second opinion.
Citations answer "where from", not "is it right"
The most substantive engineering in this release is provenance. Because the data is indexed by OpenAI rather than scraped at query time, the system can offer granular citations: a figure traces back to the page it came from, and a professional can open the evidence.
That is a real improvement, and it kills the worst failure of general-purpose chatbots in finance, which is the invented number that looks plausible.
It does not, however, do the thing people will assume it does. A citation establishes where a number came from. It says nothing about whether that number answers the question that was asked. A correctly cited figure can still be the wrong period, the wrong entity in the group, a pre-restatement version, or a definition that does not match the one used elsewhere in the same deck.
Provenance turns "is this made up?" into "is this the right one?" - a better question, and still a question a person has to answer.
What the benchmark does not cover
Three gaps are worth stating plainly, and OpenAI is reasonably candid about the third.
The evaluation methodology has not been released in enough detail for anyone outside to reproduce the score. Connector error rates are cited, but the composition of the test set behind them is not disclosed.
And the documents are one genre. Treasury filings are structured, public, and written to be read. A credit agreement, a data room in a live transaction, a management pack assembled in Excel by someone who left the firm - those are different problems, and a score on one does not transfer to the others.
OpenAI's own framing is that institutions have to validate reliability against their own workloads. That sentence is the most useful one in the entire launch, and it is the one least likely to survive into the internal memo that summarises it.
The genuinely new part is the plumbing
If you strip out the model, this release is mostly a data deal, and that is the part that changes behaviour.
Until now, most enterprise AI failures in finance were not reasoning failures. They were access failures: the connector timed out, the entitlement did not cover that dataset, the rate limit hit halfway through, the auth token expired overnight. Teams spent their time on retrieval plumbing rather than on the analysis.
Bundling and pre-indexing the data removes a whole class of those failures. What is left behind is a cleaner and harder problem: the data is right there, and the answer is still wrong about three times in ten.
It also creates a dependency worth naming. Data access, indexing and the audit trail now route through one vendor. That is a reasonable trade for speed, but it is a trade, and the people approving it should know they are making it.
How we check our own model, including where it broke
We run a public forecast experiment. Every forecast is written out in full, hashed, and anchored to the Bitcoin blockchain before it is published, so nothing can be quietly edited afterwards. Then it gets scored against the outcome, wins and losses alike.
As of 11 September 2026 the record stands at 454 scored forecasts across 39 distinct days, with 558 sealed in total. The core 50% zone has held 57% of outcomes against a target of 50%, the wide 80% band has held 89%, and the median error is 2.20%.
Read that pooled line and you would conclude the method works.
Split it by horizon and it falls apart in one place:
| Horizon | Core 50% coverage | Mean error | | --- | --- | --- | | 24 hours | 74% | +0.6% | | 7 days | 74% | +4.5% | | 30 days | 4% | +30.0% |
The 30-day band did not merely underperform. Its core zone should hold about half the outcomes and held about a twentieth, and every single miss fell the same way - the market rose further than the band allowed for. That is a method failure on a trending regime, not bad luck, and the horizon is being reworked.
We are not citing this to look rigorous. We are citing it because the failure was completely invisible in the headline number and only appeared when the results were split. A single aggregate score is the most effective way to hide the one segment where a method is broken.
Which is precisely the question to ask about 69.9% on one genre of document.
A checking step you can actually run
If someone puts this on a desk, the validation does not need to be elaborate. It needs to exist.
1. Build a set of 50 questions from your own documents, where a person already knows the correct answer and can defend it. 2. Score the system once with citations hidden. This measures the answer. 3. Score again with citations visible, and count how many wrong answers a reviewer caught by opening the source. This measures whether your review process actually works. 4. Break the results down by question type - extraction, calculation, definition, comparison across periods. A pooled score tells you nothing about where it fails. 5. Record the direction of the errors, not just the rate. Errors that all lean one way are a method problem and will not average out. 6. Re-run it monthly. The model gets updated, the documents change, and a score from March is not evidence about September.
That is roughly a day of work, and it converts a vendor's benchmark into a number about your own workload.
What to watch next
Three things would change the picture. Whether OpenAI publishes the evaluation methodology in reproducible detail. Whether any third party independently reproduces the OfficeQA Pro result. And whether pricing and eligibility become public, because a product sold through account teams to undisclosed buyers is difficult to assess from outside.
The honest summary is short. This is the strongest published score anyone has on that test, the data plumbing behind it is a real advance, and it is still a rate you have to build a checking step around before it touches anything that goes to a client.
Sources: OpenAI's announcement of ChatGPT for Financial Services (10 September 2026), VentureBeat's write-up of the data sources and OfficeQA Pro scores, and CNBC's report on the junior banker workflows it targets. Our own figures are read live from the public experiment page.
Educational content - not financial advice.