Reported by Xinzhiyuan
GPT-6 Astra has only just become fully available, and X is already being flooded with posts about it.
Scrolling through the timeline, the scene looks something like this:
• In the middle of the night, someone heard voices in their living room. They went out to look — it was NPCs they had created, holding a meeting.
• Someone jokingly asked it to make Panzer Dragoon. It took 24 minutes.
• Someone asked it to rig bones in Blender. One prompt and it was done. The poster immediately announced: “We’re finished.”
• And an immunologist who has studied T cells for 35 years used it to resurrect the first computer he owned 40 years ago. The next day, he declared: AGI is here.
In short, basically everyone who has tried it is instantly blown away.
What’s even more striking is that OpenAI insiders suddenly leaked a hint: this thing is just an appetizer.
Researcher roon just posted: “I haven’t even scratched the surface of what Astra can do.”
That alone was scary enough. Then he immediately followed up with another stab:
“I suspect it will become obsolete in some way within weeks.”
The top reaction in the comments: So Astra is already the interim version?
Okay, then.
Let’s take a closer look at what this “interim version” has done to scare everyone so badly.
— 24 Minutes to Make Panzer Dragoon - Users Say “We’re Finished” —
Japanese netizen “Yasei no Otoko” jokingly told Astra Pro: Use WebGL to make the first level of Panzer Dragoon.
24 minutes later, a playable level appeared.
You ride a dragon low over the water, fly through one broken stone arch after another, lock on, fire, and enemy airships press in from the distance. The old Sega Saturn feeling is all there.
Absolutely insane!
Comedian Duncan Trussell asked it to build “The Backrooms” in Blender and then add a VHS tape effect on top.
It took 30 minutes to build and 20 minutes to export. The sound effects were also made by the model — about five prompts in total.
Trussell was careful to emphasize: this is not AI-generated video, it’s all Blender.
Then he tried rigging bones.
A few hours later, he discovered it could also lip-sync the bones.
Upload an audio clip, and the character in Blender starts lip-syncing.
Alexey Fateev wanted a browser version of God of War.
Astra gave him “a cathedral above a dead star,” with demons swarming and a greatsword cleaving off limbs. The first version didn’t work, so he kept pushing through several rounds. Codex rewrote the whole animation system:
Feet now plant on the ground, and both hands follow through with the greatsword swing.
“Codex is the last game I ever bought.”
A playable hack-and-slash game written in Three.js.
Wharton professor Ethan Mollick threw the 1977 text adventure Zork at it and asked for a complete 3D action-adventure version.
The story and puzzles were all preserved, combat was added, and the characters and scenes were built directly with Three.js. It ran for 4 hours and 31 minutes, and the result was open-sourced.
The full prompt is in the image below (scroll up/down to view).
Peter Gostev made “Gogh Strike”: a post-impressionist 5v5 first-person shooter. The entire map feels like stepping inside a Van Gogh painting, and the painter and soldier models were rendered in Blender.
It ran on the Astra Ultra tier for 6 hours, was open-sourced under MIT, and was also turned into a ChatGPT Site.
Van Gogh-style Counter-Strike: the painter and soldier models were rendered in Blender.
Alex Finn finished a top-down multiplayer extraction shooter in a single pass, using Blender plus Unity — graphics, sound, and gameplay all included.
Faster than 5.6 Sol, fewer credits, much higher quality, no trade-offs. “This is the holy grail.”
Rogue used the xhigh tier for 40 minutes to produce a 2.5D Japanese-style fantasy map.
Cherry blossom trees, wooden houses, torii gates, and a stone bridge, viewed from a 45-degree overhead angle, with quest dialogue, a boss, and fully original assets.
It took just one sentence of prompting plus a reference image to set the style.
zcw used xhigh for an hour to clone War Thunder — “super simple prompts, almost no guidance” — and you can fly a plane.
AiBattle ran 3D Sonic in Godot: tropical coastline, checkerboard cliffs, giant loops. Using the same prompt, the Max tier took 53 minutes and the Medium tier took 25 minutes.
Max tier vs. Medium tier, same prompt.
Adam Lyttle used the Ultra tier for 26 minutes to generate a Three.js subway station scene — tiled floors, ticket gates, and a parked train. Commenters said the subway’s level of detail is unreal to the point of absurdity.
Then he burned through 10,460,032 tokens!
And then there’s playing games itself.
With Pokémon FireRed, GPT-5.5 ran for 218 hours without finishing. 5.6 Sol needed 96 hours and 35 minutes. Astra’s high tier topped it in 18 hours and 12 minutes. It relied entirely on screenshots — no memory info, no walkthrough.
Notice the third line of that screenshot: Astra’s low tier still only took 21 hours.
— Resurrecting a 40-Year-Old Computer: An Immunologist Declares AGI Has Arrived —
Games are fun, sure. But what follows is what truly makes people unable to sit still.
Derya Unutmaz, an immunologist who has studied T cells for 35 years, tried it for a few days and then directly declared: AGI is here.
He used Astra plus Blender to recreate his first computer from more than 40 years ago: the Sinclair ZX Spectrum 48K.
The rubber keyboard moves. The cassette drive makes the tape-loading sound. The built-in Sinclair BASIC can actually write and run programs. It comes with a playable old game. Even the Sinclair copyright notice on startup is there.
“The machine that led me into the magic of programming, resurrected by AI 40 years later.”
Jay Chooi had Astra and Fable 5.1 alternately control a robotic arm.
On tasks like “put the blocks into the bowl,” Astra’s completion rate was 95%. Fable 5.1’s was 40%. Astra used 6.2 times fewer output tokens and cost 2.3 times less. All 120 trajectories were made public.
Their conclusion: at this rate, LLMs may be able to control robotic arms in real time as early as the end of this year.
The design world is no different.
The founder of Adam CAD announced: GPT-6 Astra is the new SOTA for agentic CAD — a true step change in AI capability for mechanical design.
Angel asked it to draw an SVG, and out came a PS4 controller. Shadows, buttons, and touchpad texture are all there. You would never guess it was drawn with code.
This is SVG, drawn with code.
Angel also asked it to do something cool with redstone in Minecraft. It built a huge circular building — a ring of teal and purple blocks with a core embedded in the center. When powered, it spins and plays music. It named itself HELIOS.
Zachi asked it to draw him inside Canva.
He placed his own photo in the corner, and Astra painted from the background upward, stroke by stroke, on Canva’s canvas.
First came the sky and trees, then the white T-shirt, arms, and face. In the end, a realistic portrait emerged, almost identical to the photo.
Designer Marvin Schwaibold used no design skills at all. He started by setting padding, layout constraints, and a basic grid, then got two or three components working — “and then it clicked.”
“GPT-6 Astra is genuinely good at UI.”
Justin paired it with Devin and two custom skills. The prompt was just one sentence: “Give a luxury car dealer a complete landing page — understated luxury, unconventional layout.”
It came out in one pass.
Music wasn’t spared either. Someone tested it directly with Bach.
On music blogger Auggie’s Bach Benchmark, Astra’s four-part chorale had no voice-leading errors anywhere, the harmony included a Neapolitan sixth chord, and it was still “the first model to write a passing tone.”
On the video side, BURAK put Astra and Fable 5.1 side by side on the same prompt: the same ink-wash monk riding a dragon. Astra is on the left, Fable is on the right.
His verdict: Astra feels more aggressive — the camera movement, transitions, and action all feel like someone is directing the scene.
From robotic arms to Bach, from CAD to redstone — each one alone is an entire specialized industry.
And it swept through all of them in two days.
— Watching It Click the Mouse Is a Delight —
After seeing all of this, you may have already noticed:
Astra has been using the computer the whole time.
Blender, Unity, Godot, Unreal, Canva, Ableton, Devin, Minecraft — it clicks the mouse, drags models, rigs bones, and switches software.
After using it for a day, Baoyu wrote a dedicated analysis.
His feeling is that 5.6 Sol always needs to “wait a moment” before operating the computer. Astra has no such pause. It tests apps quickly and accurately. “Watching it click next to you really is a delight.”
Previously, after development you still had to manually verify the results, or rely on end-to-end frameworks like Playwright. Computer Use was never impossible — it was expensive, slow, and inaccurate. Astra has completely broken through that threshold.
From now on, when you assign GPT-6 a task, you can just let it verify the final result with Computer Use, saving most of your manual testing.
“Watching it click next to you really is a delight.”
The video at the end of that post was the one posted by OpenAI employee Victor Nunez:
“Hey Astra, how fast can you actually use a computer?”
That feeling is confirmed by the benchmarks.
Codex lead Tibo shared the Terminal-Bench 4.0 leaderboard this morning.
The result: Astra (max) ranks first at 58.2%. Fable 5.1 (max) ranks second at 57.9%. The gap is small, but the cost is only half.
Alex Finn offered a usage summary called “reverse prompting”: give enough context, give fewer instructions, and ask the model more questions.
“It’s stupid for us to teach the model how to work — it’s like me teaching Michael Jordan how to shoot.”
— So, What Is Bel? —
Back to roon’s words.
Within 48 hours of release, it had already done all of the above.
Brockman said at the launch event, “Welcome to the AGI era.” Altman has said before that by the end of 2026, there will be a model he is willing to call AGI.
And roon says this thing will become obsolete in weeks.
Blogger Haider connected that statement to a rumor from a month ago:
On August 25, well-known leaker leo claimed that OpenAI has already completed its next pretraining run, codenamed Bel — successor to Doug — with total parameters exceeding 10 trillion, on the same scale as GPT-4.5.
But whatever Bel is called, and whenever it arrives, roon’s words themselves are already a signal —
OpenAI still has cards in hand, and they believe those cards are much bigger than this one.
The current Astra can already fix a server while you’re on a date, let an immunologist resurrect a 40-year-old computer, and make NPCs hold meetings in your living room.
The one coming at the end of the year is too frightening to imagine.
References:
https://x.com/tszzl/status/2095630749403619499
https://x.com/haider1/status/2095806732341764591
https://x.com/synthwavedd/status/2092326145270456377
https://somethingbig.ai/astra-review
https://x.com/yasei_no_otoko/status/2095975195312029889
https://x.com/duncantrussell/status/2096003511104508411
https://x.com/superalesha/status/2095988972879335792
https://x.com/DeryaTR_/status/2096062355692048605
https://x.com/chooi_jeq/status/2096064315115839904
https://x.com/Clad3815/status/2095596013168050551
Editor: Moses