by Tan Aik Keong (AK)
Over the past few days, one of the hottest topics in tech circles has been GPT-6 Astra.
What's grabbing attention isn't just that it can write essays and code — it's a string of demos that feel almost hard to believe. Give it a few photos, and it can build a real, continuously editable 3D model in Blender. OpenAI even showed GPT-6 Astra building a house in Blender, then bringing it into Unreal Engine 5 as a space a person can actually walk through and explore.
That got me thinking about a bigger question than "how smart is GPT-6" — how far are we, really, from AGI?
What OpenAI actually means by AGI
AGI — Artificial General Intelligence — is usually translated into Chinese as "通用人工智能." OpenAI's own definition is worth sitting with: not a machine that scores full marks on a test, but a system that is highly autonomous and outperforms humans at most economically valuable work.
Measured against that bar, GPT-6 starts to look interesting in a different way.
From "answer" to "action"
The old ChatGPT was, at its core, a "you ask, it answers" AI. Ask it to draft an email, it writes one. Ask it to analyse a spreadsheet, it tells you the answer. Ask it to write code, it hands you the code.
GPT-6 is stepping into a different phase — it doesn't just give you an answer, it uses a computer itself to get the work done. It can open a browser, fill in forms, operate a CRM, manage a calendar, do online research, produce documents, build spreadsheets and slide decks, put together a website, and then test the front end itself. OpenAI positions it as a model capable of handling complex, multi-step professional work.
That's remarkably close to what we used to imagine as a "digital employee."
Today, I tell an AI: "Research this company, analyse its financials for the last three years, identify five risks, and put together an investment committee briefing."
In the near future, the AI might not just tell me what to do — it might actually open the browser, find the data, analyse it, build the model, produce the slide deck, and hand back a finished result.
Once AI moves from Answer to Action, the AGI conversation stops being theoretical.
Spatial intelligence, not just a 3D party trick
When people see the 3D demos, the first reaction is often: "Are 3D designers about to be out of a job?"
I think that misses the real point. What matters isn't Blender — it's spatial intelligence.
Large language models have, until now, lived in a world built mostly of text. A model knows what the word "chair" means without necessarily understanding how a chair is actually constructed in three-dimensional space.
GPT-6 can look at photos from multiple angles and reconstruct a 3D object through CAD code. In OpenAI's own BenchCAD test, it achieved 95.9% geometric overlap accuracy. It can even construct a building, then turn that building into a walkable 3D environment.
What's happening here matters: AI is working its way, step by step, from Language → Image → Video → 3D → Environment, building an understanding of the physical world along the way.
Why does that connect to AGI? Because human intelligence has never been purely linguistic. A three-year-old can't write an essay, but knows a cup breaks when it falls, a chair can be sat on, and there's space behind a closed door. Genuine general intelligence has to understand objects, space, cause and effect, and what happens after you act.
So I don't read the 3D breakthrough as "a design tool got an upgrade." It looks more like one more piece of the puzzle on AI's path toward a world model.
The benchmark that concerns me more
Another number worth watching is ARC-AGI-3 — a test specifically designed to see whether an AI, dropped into an unfamiliar environment, can observe the rules, learn, and solve new problems the way a person would. GPT-6 Astra scored 99.9%; ARC Prize also reported it hit human-level action efficiency on 96% of levels.
Even so, I wouldn't say "AGI has officially arrived." Matching human performance on a test isn't the same as possessing the full range of human intelligence. Real AGI still requires long-term memory, real-world understanding, continuous learning, reliability, and autonomous judgment — plus a thornier problem: as AI is handed more and more autonomy, can we still control it?
GPT-6 has, in fact, already become the first widely deployed OpenAI model to reach their "Critical" cybersecurity capability tier. Given the right tools and permissions, it can already discover unknown security vulnerabilities and develop new attack methods.
The line is getting blurry
So I think the real significance of GPT-6 isn't that some benchmark ticked up another few percentage points. It's that we used to talk about AGI as something ten or twenty years away. Today, that line is starting to blur.
AI can now see, hear, speak, reason, write code, use a computer, operate software, and is beginning to understand 3D space — while independently executing dozens or hundreds of steps toward a goal.
Looking back one day, we probably won't find a single headline that announced "AGI was born today." AGI is more likely to arrive the way the internet did — quietly, piece by piece, until one day we realise the future we'd been waiting for had already gotten here.
Part of the AK AI Corner column. Originally published in Oriental Daily (东方日报) on Sep 8, 2026.