4x faster than the human with Excel
OpenAI just released GPT-6 Astra, and it’s scaring a lot of people.
They gave the new model a challenge from the Financial Modeling World Cup*.
It finished about 4 times faster than the human who WON the competition.
I already thought since a few months that AI was better than me in Excel. But when I saw this, I came out of my holidays to check!
Plus, it’s not just Excel where GPT-6 excels 😉
So today I want to show you six things GPT-6 Astra can do. Plus my answer to the question most finance pros are worrying about.
Is this the model that replaces us?
OpenAI's president: “Welcome to the AGI Era”
|
|
Greg Brockman called Astra a generational leap. Then he told reporters he personally believes OpenAI has reached AGI (Artificial General Intelligence), and left it to users to decide whether this particular model is the one. He closed the briefing by welcoming everybody to the AGI era.
I don’t believe we have reached AGI (plus people can’t even agree how to define AGI).
For me, this will not be something that just appears. We won’t have the current AI one day, and then AI that replaces humans then next. It will be step by step, area by area.
Researchers like Marcus Hutter define AGI (or universal AI) through an agent’s ability to achieve goals across wide, changing environments.
So, instead of needing a rewrite or retraining for new challenges, the software adapts, learns, and transfers knowledge between completely different domains (by itself).
We are still not here.
I think that there are things where AI surpasses humans, but still a lot of things where AI does not.
And right now, AI cannot copy the diverse set of skills a finance pro needs to deploy every day to do their job.
But models like GPT-6 are definitely getting better at helping with work that does not add value, so you can spend more time on the things that do.
So let me give you the six practical ways that GPT-6 will help you (plus a word of warning).
The 6 Main GPT-6 Changes
1. The biggest change: Computer Use
|
|
Computer use is basically AI controlling your mouse and keyboard. Seeing what you see on screen, and carrying out actions like you would.
It opens applications, clicks, types, moves around interfaces, works with your files, and then self-reviews.
This might remove a big bottleneck for you. You can’t connect all of your finance tools to AI, so a lot of your processes still need somebody sitting there moving between Excel, a browser, a BI tool, an ERP screen and PowerPoint. Previously this was you, and it caused a lot of headaches (and sometimes human error).
Some examples;
- In Excel: The difference is execution. Before, you asked for a formula and you pasted it in yourself. Now it navigates a complex workbook, understands the task, changes the model, checks the output and keeps going. That is closer to an AI analyst operating the spreadsheet than an AI assistant sitting next to it.
- In Power BI: The old prompt was "give me the DAX and the instructions to build this dashboard." The new prompt can be "build a CFO dashboard comparing actual against budget, explain the largest variances, improve the visual hierarchy, validate the calculations and prepare it for management review."
- Form filling: Imagine the process of having to fill out slightly different vendor forms. They have the same information, but because they all have different layouts, you would have no option but to do this manually. Not any more.
- And at month-end: Imagine the whole run as one job. Open the files, update the model, investigate the variances, make the charts, refresh Power BI, populate the management deck, draft the commentary. It chains this steps instead of stopping after each one and waiting for you.
2. Fewer wrong numbers
On OpenAI's own hallucination benchmark, Astra sits at 4.2% against 12.2% for the previous model. On top, they say it is three times less likely to claim it did something it did not do. For us in finance this is super important.
3. It asks before it guesses
OpenAI put the two models on the same job side by side. The previous model built an entire career site in 13 minutes 15 seconds without asking one question. Astra stopped after 20 seconds to ask what career the person was moving into.
4. A million tokens of context
The context window is around 1,050,000 tokens. Your accounting policy, your account mapping and the dataset can all sit in the same conversation without you losing the quality of the conversation.
5. It stays on one long job
Long multi-step work is better now, so a monthly review can be one connected assignment instead of fifteen separate chats where you re-explain the context every time.
6. You can correct it mid-task
You do not have to restart when something is wrong. Send a short correction tied to the calculation in front of you, like telling it to treat OPEX as positive but display it as a deduction, and it recalculates the dependent outputs.
272,000 tokens limit (important to know but nobody talks abou it)
|
|
Warning: cross 272,000 tokens and your bill doubles mid-request
Astra costs $10 per million input tokens and $50 per million output. That is exactly 2.5 times the previous model at $4 and $20, and OpenAI says that lower price is promotional at least until 21 November.
Now, go over 272,000 tokens of input and the whole request reprices to $20 in and $75 out. Attach a year of GL detail to a variance question and you might end up doing this by accident.
Now these costs are more for companies using GPT-6 for developer work using the API.
But, for you, what it means is that you might end up reaching your usage super fast.
(some people say they have consumed their usage in one prompt).
So the practical advices here are boring ones.
Only use GPT-6 for your most (really) complex tasks.
The one thing to remember
Everybody wants me to say either it changes everything, either it changes nothing.
It is neither.
Think of it like this. You get a better iPhone, you get excited but when using it: it doesn't feel different. But for some advanced uses, some changes are a game changer: for example now the camera is so good that most social media short videos of big influencers are recorded with iPhones and not professional cameras.
GPT-6 is the same (I also said this about Claude Fable). It will not change anything for a normal finance person who does not use the functionalities that matter (Projects, Work, Skills, ChatGPT in Excel and Powerpoint).
So my two most important advices has not changed (and these advices, nobody are giving them to you besides me):
1) Get good at using the full capability of any tool. Then send 90% of your finance tasks to the less intelligent models to save your consumption, and keep 10% for a model like GPT-6.
2) Don't switch from Claude to ChatGPT just before OpenAI made a super cool marketing about how good ChatGPT-6 Astra is, instead get the best out of your current tool and in a few weeks, you will see that another AI company will come up with a better model anyway 😉
So the good news: today, nothing has changed for you. But in six months a lot changes for the people who start using more and more AI and if you don't know where to start, join me tomorrow here.
You haven't been replaced yet 😉
Best,
Your AI Finance Expert,
– Nicolas
P.S – The best way to learn from me directly is to join my weekly masterclasses. 60-mins + Q&A where I’ll get all your AI in Finance questions answered. Join me here.
P.P.S. – Worried about burning through your AI usage? Here's how to use 97% less without losing quality → How To Save 97% of Your Claude Tokens As a CFO
*To find the Financial Modelling World Cup test. Go to the OpenAI GPT-6 Astra release page. Scroll down, and click the tab for 'Excel Competition'.