It's the first time in the history of the world.
The AGI's fire is still raging, and the next contest has begun!
When the GPT-6 Astra appeared, it directly blew up all the major lists, with no match.
Just today, the most recent MazeBench benchmark for 3D space reasoning, Astra, which has run more than 60 hours, dropped 14%, to number one.
After all, OpenAI is the first team in the world to take over AGI.
Today, not everyone has eased the shock of the GPT-6 Astra, and a new wave of cases is emerging.
Build a cell model, watch a video of the design piece, identify the sound from the spectrum...
It's just a "monster" of 100,000 GPUs, and it's hard to imagine how much the next generation of GPUs will be against the clock if 400,000 GPUs go down.
GPT-6 slaughterhouse space reasoningThe highest score of 14%.The MazeBench test is a 3D world created by Jonathan Pappas and David Pappas in late July.
The world overlooks MazeBench, and the little yellow is the smart one.
In total, it has more than 200 rooms, containing 100 gems, one point, and a full 100 points.
The room is full of box-shaped puzzles, extra elevator racks, switchable walls, sliding ice surfaces. It falls into a deep pit and can only be removed, repositioned, or transported back to the room where it was.
Green is the player, blue is the box.
For example, at this point, the blue box is pushed into a black hole, filling a bridge to pass through.
The camera can rotate up and down and look at the same scene from 20 angles.
Left: 3D view; right: text the model actually sees
The jewels are hidden in this place.
The model received only 11 moves through the MCP, four mobile, four rotary cameras, plus cancellation, reset, and transmission.
The difficulty is that a problem often involves planning more than 100 steps. And it's real 3D, with the same wall looking up and down and looking from the side.
As a result, as soon as it was released at the end of July, the best of the smart bodies didn't make it out of the first few rounds.
At the time, using Python's GPT-5.6 Sol took 13%, Fable 5 11%, and Opus 5 9%.
And then Astra came along -- not open Python, just take 14% off.
It can be said that in 3D space logic, GPT-6 Astra and all previous Agents are no longer at the same stage.
Below is an exploratory heatmap of it and Fable 5.1. Fable 5.1 stops for a little bit around the corner, and Astra almost flips the entire map over again.
In order to save Token, the test team changed the GPT-6 to a "batch invitation" instead of taking it one step at a time.
As a result, it can look 10 to 20 more steps forward each time, directly cutting the 3 billion tokens it was supposed to burn down to 350 million.
In the Huangdaway test, this batch model makes the cost of Astra two-thirds cheaper, but the cost is 60 percent longer thinking time.
Speaking back, however, Astra really wants to break MazeBench's hard bones, and there's a long way to go.
ARC-AGI 3 has been upgraded to 99.9%, but on the MazeBench, the strongest AI is only upgraded to 14 gemstones.
Because, in the 3D understanding of space, Astra has a biased phenomenon.
It prefers to look at the gates from a vantage point, as a result of which walls that can only be seen from a 3D perspective are directly transformed into air walls.
Newcomers teach to turn it off, but once it hits a 3D puzzle that is over three layers high, it immediately catches its eye.
In addition to brushing the puzzle, the full network developers have put the actual task in the hands of the GPT-6 Astra, playing a new wave of flowers.
Astra's new "open mode"It exploded.Just this morning, Greg, President of OpenAI, released several surprising demos of the GPT-6 Astra.
Each one is not a trick, it is "Astra's way of solving real problems".
The showered hair, it settled.In the words of CEO Daniel Griesser, "Yesterday, I really felt the AGI!"
He made the GPT-6 Astra do something that was actually ground gas.
Daniel recorded a four-minute video in which an old man's hair clogged his spit leaked out of the bathroom, and he got so pissed off.
He weighed the parts in front of the camera with a numeric meter, and then turned the video over to Astra Ultra: give me a trick.
Astra's next move was to dismantle the video, while listening to his voice-description, and to align the movements of each tonne of dimensions.
In the end, Daniel not only gave a design to clean his hair easily, but also pioneered an interactive sketch for him to confirm.
Confirming the sketch, Daniel directly commanded Astra to take over the computer with "computer use" and open his own Fusion 360 drawing.
The other person sat by and watched with tears as AI built the parts one step at a time, and it was a fully parameterized model.
By the way, I've also stolen a few Fusion hidden tricks.
When he finished building the model, he said "add a reverse angle" and threw it into the 3D printer.
In response, Daniel said, "It's not a high-tech thing, but it does solve a big problem".
The most important thing is that you're too fast to draw.
Scroll down and see
The developer, Andrew Aiginin, also exclaimed, "GPT-6 Astra is incredibly powerful".
The day before, he had just asked Astra to build a cell model, and the day after, he had it explain how Ozempic works.
An hour later, Astra delivers an interactive molecular journey of 11 chapters, from the GLP-1 receptor binding to the amplification of the cAMP signal, all the way to the insulin stage.
In the video, the right-hand side is a 3D molecular model that can rotate and scale at will, while the left-hand side shows the number of active receptors in real time.
A spectrum map to identify soundsWhat's more, the GPT-6 Astra is also capable of "seeing guesses".
Networker Max did a highly modified test that gave Astra a mailing spectrum map, without audio, and directly copied the mailing.
Try "listening", what's that sound?
 |  |
Astra looked at it and decided that it was most likely that the dog was barking repeatedly, even with a slightly rough lower jaw.
What's more, the rhythm of its screaming makes the brain full of woof-woof, woof... woof, woof-woof.
Moreover, it also explains that the head is a path, with a strong low-frequency subject, similar resonance bands, and a noise-induced high-frequency edge, and that repeated resonance patterns are more like barking than talking.
It's not a sample, and it's only light reasoning.
In exchange for another diagram, Astra guessed better, directly giving a series of images: lighted sword fires, low-frequency flashes, and finally extinguishing the sword.
Max couldn't bear to feel this, and what we've dug up now, I'm afraid, doesn't even count the skins of this model's ability.
48 check codes, one pass.In addition to speculation, Astra also opened the door directly to humans and machines.
There's a game called "I'm Not a Robot" on neal.fun, with 48-bit verification codes, one at a time for drilling, picking out dolls or pancakes, ordering fire extinguishers, and numbering in an embedded grid, which many people don't go through themselves.
Sharif Shameem allowed Astra to play and the result was complete.
The result was not only that the surrounding netizens shouted, "Verification codes are dead".
The game's original creator, Neal Agarwal, also blew it up.
He could not help but say, "Go on, I've done my best. Good luck to you".
In addition to that, there are others who play wild.
Astra uses Blender to recreate Michael Jackson's classic dance moves.
After this demo, I can't help but wonder: how we opened up to Astra is really too conservative.
AGI is here.There's only one ASI left.The GPT-6 Astra wave was born in space, making it impossible for old Yellow to cry out "AGI is here!"
But honestly, it doesn't make much sense to argue about whether Astra is "true AGI" anymore.
AI itself writes code, trains the next generation of AI, AI research trainees are officially inducted, and the next generation of OpenAI, Bel, Anthropic Mystery Model 2, is also in the pipeline.
The coordinate axis of the entire AI competition is changing dramatically
The focus has shifted from "who gets to AGI first" to "who can accelerate the intelligent self".
At the end of the night, there was only one ASI remaining.
And the scales made by human hands are being dismantled by new species one by one.
In the face of the collapse of the old system, we published our own "ASI coordinates" on Aiera.com.cn.
Today's panel focuses on the latest models released in the past seven days, which are updated daily in the morning.
The latest ranking has been revealed: GPT-6 Astra led by a high score of 90.3, Fable 5.1 by 84.0, and Gemini 3.8 Flash by 73.8.
The week's overall chart, however, overwhelmed the model with bulk data.
The first was Claude Opus 5 at 89.2, the second was GPT-5.6 Sol Xhigh at 89.0, and the third was Claude Fable 5 at 86.2.
Astra has not yet settled the weekly list, and until Wednesday, when the matrix is reconstructed, it will claim the new throne.
This is what the new ASI coordinate system is really going to measure: AGI is a power line and RSI is an acceleration curve.
The old scale, which could not be measured in the new century, was the same as the old scale.
AGI's horizons have crossed.ASI, we are in the final city pool, waiting for you.
Read more: Just recently, the new website was born!ASI countdown time started