GPT-5.5 Is Out. Here's What Finance Teams Should Actually Test First.
GPT-5.5 Is Out. Here's What Finance Teams Should Actually Test First.
Labels: AI Finance
OpenAI released GPT-5.5 on 23 April — exactly one week ago. The launch cycle for frontier AI models is now moving fast enough that a new release can feel routine before most people have had a chance to test it properly. So I spent the last week actually running it through the kinds of tasks that matter in a finance context, rather than reading benchmarks and forming an opinion from a distance.
This is an honest account of what I tested, what impressed me, and where I think the hype gets ahead of the reality. The goal isn't a comprehensive review — it's to give finance teams a practical starting point so you're not wasting hours testing the wrong things.
There are enough genuine improvements in this release to warrant the attention. But the way it matters for finance is more specific than the marketing suggests.
|
84.9%
GPT-5.5 score on GDPval — a benchmark testing real-world professional knowledge work across 44 occupations. Finance is one of them.
|
88.5%
Score on OpenAI's internal investment banking modelling benchmark — the highest finance-specific performance figure OpenAI has published for this model.
|
|
2 weeks
Time saved by OpenAI's own finance team using GPT-5.5 in Codex to process 24,771 K-1 tax forms across 71,637 pages — a task previously consuming significant analyst hours.
|
40%
Fewer tokens used compared to GPT-5.4 — which means faster responses and lower API cost for teams running it at volume.
|
What's Actually Different for Finance
GPT-5.5's headline improvements are in agentic capability — the model's ability to take a messy, multi-part task, figure out what needs to happen, and work through it with less step-by-step guidance from the user. OpenAI president Greg Brockman described it as a model that "can look at an unclear problem and figure out what needs to happen next." That's a meaningful shift for finance applications, because financial tasks are rarely clean single-step requests.
Two other changes matter for finance use cases specifically. First, the model uses significantly fewer tokens than its predecessor to produce equivalent-quality outputs — which matters both for cost and for response speed in higher-volume workflows. Second, early enterprise adopters have noted notably lower hallucination rates, particularly on structured data tasks. The CIO of Bank of New York, which was among the first institutions to test the model, described a "step change" in accuracy that is directly relevant to financial work where precision is non-negotiable.
What I Actually Tested
Multi-document analysis — contracts and agreements
Finance teams routinely deal with a stack of supplier agreements, funding contracts, or service arrangements where the critical terms — payment conditions, indexation clauses, termination provisions — are buried across dozens of pages in different formats. Extracting and comparing those terms manually is tedious and error-prone.
GPT-5.5's 1 million token context window means it can genuinely hold multiple long documents in a single session. The model handled cross-document clause comparison well — identifying where two agreements had inconsistent payment terms, for example — with a level of accuracy that made it useful rather than just indicative. For first-pass due diligence and contract review in finance, this is a real time-saver. The final review still needs a human. But the prep work is significantly faster.
✅ Strong — practical time saving for document-heavy finance tasksData analysis from messy, multi-source inputs
I tested feeding in unstructured financial data — a mix of exported figures, narrative context, and prior-period comparisons — and asking the model to identify drivers of variance, flag anomalies, and summarise the picture. The task that previously required a finance analyst to manually piece together from multiple sources.
The results were genuinely useful. The model was good at identifying patterns across messy inputs and surfacing the questions worth investigating. Where it struggled was in distinguishing between a variance that was explainable from context and one that required external knowledge to interpret — the "that number is wrong because the payroll system had an error in week three" type of context that only a finance professional with operational visibility can provide. The analysis was a strong starting point. It was not a finished product.
✅ Good — strong starting point, human interpretation still requiredFinancial commentary drafting from raw figures
I provided period-end figures, prior-period comparisons, and relevant context, and asked for a first draft of management reporting pack commentary. The draft quality was notably better than earlier models — appropriate language, logical structure, and the right movements leading. Where it falls short is applying organisational context that wasn't in the prompt: the salary overspend that was a deliberate one-off, the board conversation that shifted the capital plan, the subsequent receipt that offset the revenue shortfall. The draft needs a finance professional to make it accurate. But getting to 70 per cent of the way there in minutes rather than an hour is a meaningful efficiency gain.
✅ Good — strong draft quality, organisational context must be added by humanMulti-step research workflows — regulation and policy tracking
Finance teams in NDIS, aged care and childcare spend real time tracking regulatory changes — pricing determinations, amended award rates, compliance deadlines. I tested the model on a multi-step research task: identify relevant recent changes, summarise the operational impact, flag what needs action by when. The agentic improvement was most noticeable here — the model planned its approach, worked through sources logically, and produced a useful briefing structure. The limitation is recency and primary source verification. For compliance questions with legal weight, always go to the source. For initial orientation to a regulatory change, it genuinely saves time.
⚡ Useful with caveats — verify compliance conclusions against primary sourcesWhat Still Needs Human Judgement
GPT-5.5 is better at ambiguity than its predecessors. But it still doesn't know what it doesn't know — and in finance, the most important information is often context that exists outside the documents provided. Operational decisions that shifted the numbers mid-period. Board conversations that changed the capital plan. The staff member leaving who explains the payroll spike. The model also has no feel for organisational politics, for what a particular stakeholder is sensitive to, or for when a number needs careful contextualisation rather than straight reporting. These judgement calls remain firmly in human territory — and they're the most valuable part of the job.
Where to Start
If you want to use GPT-5.5 productively, the sequence I'd suggest: document analysis first, commentary drafting second, data analysis third. These tasks have clear enough outputs that you can evaluate quality quickly, and the time savings are meaningful even in early use. Research workflows come later, once you're comfortable with the model's limitations on primary source verification. The teams that get real value from this model treat it as a capable analyst to supervise — not as a replacement for finance judgement, and not as an afternoon experiment. The use cases are genuine. So are the boundaries.
⚠️ A note on AI model training and your data: OpenAI's Business and Enterprise plans include data isolation — your inputs are not used to train future models under those configurations. The default Plus plan operates under different terms. Before using any AI model for tasks involving financial data, client information, or operationally sensitive material, verify which plan you're on and read the data handling policy. This applies to GPT-5.5 as much as it applies to any AI tool.
Ready to Build AI Into Your Finance Workflow Properly?
PFL helps finance teams move past experimentation and into structured AI adoption — identifying the right use cases, building the right controls, and measuring the time savings. If you want to make GPT-5.5 or similar tools work for your function in a way that actually holds up, let's talk.
Get in Touch with PFL →
Comments
Post a Comment