A model that drives your computer: what opens up for a small business, and what stays locked
On 3 September 2026 OpenAI released a new model called GPT-6 Astra, for ChatGPT and for Codex. The part that matters to a small business is computer use. The model fills in browser forms, moves between pages, works inside spreadsheets and runs a whole sequence of steps without a person clicking at each one. On OSWorld 2.0, the benchmark that measures exactly this, it reached 72.6% and cut roughly 47% off the time per task compared with the previous generation. On a Mac it can run that sequence in the background while you keep working.
The capability itself is not new. What changed is the success rate and the speed, and a jump that size starts to close the gap between a demo and something you can hand real work to.
Who gets it, and when
The rollout is staged. Organisations in the Daybreak program went first, the route through which approved bodies work with the strongest models. After them, within days according to OpenAI, access opens to ChatGPT Plus, Pro, Business and Enterprise subscribers, and through the API and AWS. Pro, Business and Enterprise also get a version called GPT-6 Astra Pro.
Usage counts against the allowance of the subscription you already pay for, and anyone who exhausts it can buy additional credits. On the API the price is 10 dollars per million input tokens and 50 dollars per million output tokens, with a fast mode that costs double and runs up to 2.5 times quicker. For comparison, Claude Opus 5 sits at 5 and 25 dollars. Anyone building automation on the API rather than consuming the model through a subscription should do that arithmetic first.
The part that is locked
On 1 September, two days before the launch, OpenAI published a safety update stating that Astra is its first model to cross the "Critical" cyber capability threshold under its Preparedness Framework. That threshold is defined as being able to find previously unknown security flaws and develop working exploits against real, hardened systems without a person directing each step.
That is why the rollout is slow and fenced. OpenAI reports stronger resistance to jailbreaks, stricter isolation of the model and encryption of checkpoints, and misalignment monitoring on tool-using runs. For an ordinary user the model refuses advanced cyber tasks, writing proof-of-concept exploit code among them. Anyone who needs those capabilities for legitimate security work goes through approval in the Daybreak program.
What it means for a one-person business
The work that suits a model like this is the repetitive work nobody enjoys. Pulling bank statements and arranging them into a spreadsheet. Going through a folder of receipts and lifting the supplier, amount and date out of each one. Filling in the same supplier form for the third time this month. Taking a report out of a system and turning it into a table you can actually read.
The line falls where an action commits you. An agent that fills in a form on the Tax Authority portal and presses submit has filed a report in your name. If a figure in it is wrong, the responsibility is yours, not the model's and not that of the company that sold you the subscription. The one practical rule that holds here is that the model prepares and you approve, particularly on anything touching a periodic VAT report or advance payments. Separating preparation from submission costs a few seconds and saves an amended filing later.
The second point is permissions. A model that drives a computer works with whatever is open to it: your browser, the accounts signed in there, the files it can see. The better the capability gets, the more it is worth running on a separate user with narrow access rather than on the account that holds the bank, the portal and the mailbox in one place.
The announcement about the critical cyber threshold belongs to the same picture. The capability is inside the model, and the guardrail on it is a decision made by the company running it. That is worth holding in mind before handing an automated agent the keys to systems that move money.