What happened
On 3 September 2026, OpenAI released a new AI model called GPT-6 Astra. If you went looking for “ChatGPT 6” and found nothing, that is why — the product was never given that name, which is a large part of why the launch confused people.
Where you would actually run into it: inside a paid ChatGPT plan. On the Pro, Business and Enterprise plans it shows up in the model list as GPT-6 Pro. On the cheaper Plus plan it is not in the ordinary chat window at all — you only reach it by switching ChatGPT from “Chat” to “Work” mode. Plenty of people concluded they had not been given the new model when in fact they were looking in the wrong place.
What's actually new
Most of the announcement is written for engineers. Three of the changes are worth a business owner's attention, and they all point the same way.
It can operate software, not just describe it
Older models could tell you the steps to update a record in your system. This one can go and carry them out — open the web app, click through it, fill in the form, move the data across. In OpenAI's own testing it finished that kind of task in roughly 40 minutes where the previous model needed about 75. Still slow by human standards for a single form; genuinely useful for two hundred of them.
It produces finished files, not just text in a chat box
Documents, spreadsheets and presentations that follow a template you hand it — and, the part that is actually hard, that it can rework when you change your mind halfway through.
It stays with a long job
Previous models were built around one question and one answer. This one is built to keep working on a task for hours, which is what the other two changes need in order to be worth anything.
The pattern under all three: the shift is from AI that answers to AI that does the work. Everything else in this article — where it helps, where it fails — follows from that one change.
How it helps a small or mid-sized business
The jobs that fit share a shape: repetitive, well-defined, nobody waiting on the other end, and you can tell at a glance whether the result is right.
- The supplier price list. 300 rows arrive, the price moved on 40 of them. Instead of reading it against last month's line by line, you get back only what changed.
- Descriptions for a new shipment. 40 items that each need a name, a description and a category before any of them can go on the site.
- The monthly close for your accountant. The sales export, the card statement and the invoices, into one sheet in the same format — every month, the same shape.
- A quote on your own template. Same layout every time, different scope and different prices for each client.
- Checking what competitors charge. Ten sites, the same ten products, one table — before you decide whether to move a price.
Notice what these have in common: they are all work that happens behind the business. An hour of them disappearing from your week is a real hour, and if the output is wrong you find out before anyone outside the company sees it.
What it doesn't change
This is the part the announcement is quietest about, and the part worth reading twice.
It is slow when it is thinking hard
On its highest setting it can spend several minutes working before it writes a single word. That is fine for a report you asked for over coffee. It is unusable for anyone waiting on the other end of a message.
It is expensive per job
Roughly five times the price of the model most people were using before it. Sensible for something you run a few times a week. Ruinous for something that runs on every message a customer sends.
It knows nothing about your business
Your prices, your stock, your opening hours, which supplier you actually use — none of it, unless you give it. This is the gap that surprises people most: the model can be excellent and still confidently quote a price you stopped charging two years ago.
It is still wrong sometimes, and it sounds equally sure either way
Independent testers who scored it against the previous generation found it is not broadly cleverer — it is markedly better at doing, and roughly the same at thinking. Check its output the way you would check a new employee's first week: closely, and for the first few weeks.
A long job can stop halfway
New safety monitoring can pause a task partway through and ask a person to look at it before it carries on. Worth knowing before you leave something running overnight and assume it finished.
What to do this month
If your ChatGPT plan already includes it, the whole experiment below costs you an afternoon and nothing else.
- Pick one task that eats about an hour a week and that nobody in the business enjoys. One. Not a strategy.
- Give it the real thing, not a description — last month's actual spreadsheet, your actual template. Most disappointing results come from a vague brief rather than a weak model.
- Run it twice — once on a clean example, once on the messiest one you have. The second run tells you far more than the first.
- Check the output line by line the first three times. If you cannot face checking it, that task is the wrong one to start with.
- Only then decide whether it is worth paying for. An hour a week saved is worth real money; an hour a week of checking its work is not.
Do not start with anything a customer sees. Start where a mistake costs you five minutes of embarrassment inside the office, not a customer relationship. Move outward from there, once you know how it fails.
How to read the next announcement
There will be another launch within months. Four questions get you to the answer that matters faster than any coverage of it.
- Can I actually get it, and on which plan? A large share of every launch is either unreleased or sitting behind a tier above the one you pay for.
- What can it do that the last one couldn't? Test scores almost never change your week. “It can operate software” does.
- How fast, and how much? Those two numbers alone decide whether it can go anywhere near a customer, or only near your back office.
- What did they not say? The limitations are always the shortest section in the announcement and the most useful one in it.
Most AI announcements change nothing for a small business in the month they land. That is the normal outcome, not cynicism. The ones worth your afternoon are the ones that take a specific job off your desk — and you will usually recognise them by how boring that job is.