This time it is not you asking a question and it answering with a paragraph. You tell it “get this done”, and it opens the software itself, fills in the forms, finishes the reports and hands you a finished deliverable.
On 4 September 2026, in the early hours Beijing time, OpenAI formally released its new flagship model GPT-6 Astra. Closing the media briefing, OpenAI president Brockman said something weighty — “welcome to the AGI era.”
And what I care about more is the opening line of Cailianshe's report:
The throne of the most powerful model on Earth changed hands again within two days.
Two days earlier Anthropic had just shipped Claude Fable 5.1. One night's sleep later, the throne had a new owner. The hand-to-hand combat among the giants is now counted in days.
But today I do not want to talk about parameters or who scores higher on benchmarks. I want to talk about something that genuinely hits ordinary people who make a living from a computer —
AI has crossed, for the first time, from being able to chat to being able to get things done.
1. This generation, AI learned to operate your computer
Over the past three years, what did the AI we knew actually look like?
You gave it a sentence; it gave you a paragraph. You said write me a proposal and it really wrote one. But when you opened Excel, clicked into that admin panel, logged into that system, AI could not reach — it could not read the screen and could not touch your software.
So AI automation used to depend on APIs, on MCP, on installing a separate plug for every piece of software. Before AI could do one thing, the whole software world had to be rebuilt for it.
GPT-6 Astra sets out to break exactly that.
Its headline capability is Computer Use — letting the model read the screen directly, understand interface state and operate the software we use every day with mouse and keyboard. Open spreadsheets, fill in forms, update customer records, organise schedules, build financial models, even file your taxes, build your website or design a PCB.
In OpenAI's official demo you give only a final goal; it plans the steps, switches tools and corrects itself when it makes a mistake, with no step-by-step instruction from you.
Two numbers will give you a feel for it:
- On OSWorld 2.0, the benchmark for computer operation, Astra scores 72.6%, against 65.7% for the previous generation GPT-5.6;
- Average time to complete a single computer task fell from about 75 minutes to 40 minutes, close to a doubling of efficiency.
Earlier generations could operate a computer too, just slowly enough that you would rather do it yourself. Once completing a task compresses to 40 minutes, AI computer operation has for the first time the productivity to genuinely do the work for you.
2. What is really frightening is that it can now hunt for vulnerabilities itself
If operating a computer still sounds like helping me get things done, the next item is a different matter entirely.
OpenAI says GPT-6 Astra is its first model to reach Critical cyber security capability. What does that mean? In the company's own words: given the right tools and permissions, it can autonomously find previously unknown vulnerabilities and devise exploitation plans — no longer needing a human to guide it through where to look and what to try.
We used to think AI found vulnerabilities as an assistant to human security experts; now it can be the expert itself.
That is also why it was not released all at once, but first opened to cyber defence programmes and then rolled out in stages. The stronger the capability, the more care is needed in how it reaches people's hands.
By the way, that is not the whole story. According to multiple disclosures, Astra is OpenAI's largest training project to date, completed in Texas on more than 100,000 GPUs; maximum single output is 128K tokens, with knowledge cut off in April 2026. Reasoning comes in five levels from low to max, adjustable to the weight of the task.
And the price? USD 10 per million input tokens and USD 50 per million output tokens — exactly level with Claude Fable 5. The giants have settled it: the strongest model costs this much.
3. Why I call this a watershed, not another upgrade
You can take all of the above as news, but I suggest thinking about its nature.
For three years large models have grown smarter every year, but fundamentally they did the same thing — generating content: answering questions, writing articles, writing code, drawing pictures.
Computer Use means AI is starting to move from being a generator of content to being an executor of tasks. It is not satisfied to hand you a piece of code; it wants to run the code itself. Not satisfied to give you a process recommendation; it wants to do the work inside the process itself.
Brockman said something crucial at the briefing: the AI industry has long been stuck writing a separate connector for every tool, and once computer operation matures, agents can move directly between spreadsheets, forms and web pages — enterprises no longer have to wait for the entire software world to be made AI-ready first.
Do you follow? It means that previously a company wanting AI automation had to spend heavily to rebuild its systems to be AI-friendly; now it does not. AI learns to use your old software by itself.
For developers and enterprises, the significance is comparable to what happened after the iPhone appeared, when apps no longer had to be adapted for each individual handset.
4. So what should ordinary people actually worry about?
Every time a giant ships a new model, some people panic: AI got stronger again, is my job about to go?
My view may be counter-intuitive — what you should fear is not that the job disappears but that your old way of working disappears.
What models like GPT-6 hit hardest are the skills built on being fluent at operating a pile of software, forms and systems. The Excel tricks, the reporting flows, the system manuals you spent three years mastering — AI now finishes them in 40 minutes, and without being taught, just by looking at the screen.
But the flip side is a huge window of opportunity:
In the past, one person having all of writing proposals, building spreadsheets, designing, building websites and running data was almost impossible; it took a team. Now one person plus an AI that gets things done can match a small team.
So the people who thrive in future will no longer be those most fluent in a single skill, but these two kinds:
1. People who can break work down — turning a big goal into steps an AI can execute one by one, then watching it get each step right. This will be called task design ability, and it is worth far more than fluency in one piece of software;
2. People who understand across domains — AI can finish the execution, but whether a thing should be done at all, what counts as good, and how to communicate with a client, a boss or a team still needs a human. Judgement, taste and collaboration are not things AI can replace any time soon.
In other words, AI has pushed the cost of how to do the work to the floor, so human value moves up to deciding what to do and to what standard.
5. One thing you can do starting tomorrow
Do not panic, and do not rush to learn a pile of new tools. First shift your mindset: stop treating AI as a chat box you ask questions and start using it as an employee you assign work to.
Next time you have a repetitive, clearly stepped task that needs several pieces of software — organising reimbursements, updating a customer list, writing a weekly report, or researching material and producing a summary — first ask: can I describe this goal clearly and hand it to AI, then watch the result and make the judgements?
Once, twice, three times, and as handing work over rather than doing it becomes a habit, you are gradually becoming the person who breaks work down and decides.
The throne changing every two days is the giants' business. Whether you can switch from doing the work yourself to directing AI to do it is your own business.
The sooner you take that step, the more it is worth.
#AISecondBrain #ArtificialIntelligence #Productivity #WayOfWorking #GPT6