After last night's global AI outage.
GPT-6 Astra finally made its official debut at 3:33 AM Beijing time.
If one sentence from OpenAI could summarize GPT-6 Astra,
it would probably be this:
This is the world's most intelligent and most aligned model.
And the memes have already started appearing.
This time, OpenAI has proven its strength, and it feels a bit like the GPT-4 moment—a return of the king in all aspects. What makes me happiest is that this is finally no longer a model purely specialized for coding.
GPT-6 Astra is truly a PhD-level employee with extremely strong aesthetic and professional capabilities.
The new model has many features and characteristics, and the amount of information is explosive.
But the tragedy is that OpenAI has also learned some bad habits.
Today it is only open to select organizations; it will be open to subscribers in the coming days.
Wait, brother, that's not what you said when you tricked me into buying the $200 Pro membership... You said Pro members would get priority access to new models every time...
It turns out these large model companies are all players.
I've never wanted to accelerate time so much just to use GPT-6 Astra...
However, after looking through almost all the information and materials, I think there's still a lot worth discussing with everyone.
Let's go through them one by one.
1. GPT-6 Astra Basic Information
Let's summarize the basic info first.
The model ID for GPT-6 Astra in the API is gpt-6-astra.
The context window is 1.05M, which is about 1.05 million tokens.
The maximum output length is 128K tokens.
The knowledge cutoff date is April 30, 2026.
There are five levels of reasoning intensity: low, medium, high, xhigh, and max.
The standard API price is $10 per million input tokens and $50 per million output tokens.
This price is basically identical to Claude Fable 5, but cache reading is more expensive than Claude Fable 5.1.
Here are the benchmarks; I'll discuss them in detail later, but take a quick look for now.
The total parameters are estimated to be at the 5T level. In a closed-door media briefing before the release, they mentioned that GPT-6 Astra is OpenAI's largest training run to date, using over 100,000 cards.
2. The World's Best Model for Operating Computers
GPT-6 Astra has what might be its most important positioning:
The world's best model for operating computers.
In the past, when we talked about Agents, we often talked about APIs, MCP, CLI, and so on.
The ideal state for an AI to operate software is for that software to have a dedicated interface for AI, allowing direct low-level operation. This is the most convenient.
For example, calendars have calendar APIs, email has Gmail APIs, etc. This is obviously fast.
But the problem is, the world is even more primitive and makeshift than we imagine.
In the real world, the vast majority of software might not have these things at all.
Even many internal corporate systems were written twenty years ago; forget APIs, people don't even know where the documentation is.
But if we go back to the most basic level, you'll find that the essence of all software logic is interaction. According to our current computer operation logic, that means looking at the screen, finding buttons or input boxes, clicking the mouse, entering content, and waiting for feedback.
Isn't the entire computer interaction system the best API?
So, GPT-6 Astra has massively strengthened the logic for AI to operate computers. If you don't give me an API, it doesn't matter; I'll just operate the computer myself, just like a human.
Now, Astra can directly operate Excel, Power BI, perform frontend QA, install software, test software, and look at error messages on the screen to continue troubleshooting.
The official demo even showed Astra directly operating circuit board design.
GIF
It also formats legal documents, handling title spacing and page layout.
There's a reliable evaluation here called OSWorld.
You can simply understand it as:
Throwing the AI into a real computer and letting it work on its own.
Let it open several applications, give it a task, and see if it succeeds.
GPT-6 Astra's completion rate reached 72.6%, a significant increase from GPT-5.6 Sol's 65.7%.
Furthermore, Sol took an average of about 75 minutes to complete a complex simulated task.
Astra only needs 40 minutes, about 47% less time.
ScreenSpot-Pro also jumped from Sol's 76.9% to 92.7%.
This benchmark mainly looks at whether the model can understand the screen and accurately locate where to click. It is now almost perfectly accurate.
So this positioning is quite interesting. In the future, GUI might gradually become the most universal API in the Agent era.
Any digital work that a human can complete through a screen will theoretically gradually enter the operational range of Agents.
You don't have to wait for some crappy ERP system written in 2007 to connect to MCP or open an API for you.
The AI will just look at the screen and get it done.
3. ARC-AGI-3 Reached 99.9 Points
If you don't know what ARC-AGI-3 measures, it's easy to think:
Oh, it's just another benchmark with a full score, what's so special about that?
But this thing is quite different from typical large model exams.
On March 25 this year, the ARC Prize officially launched ARC-AGI-3.
It designed hundreds of new interactive environments and thousands of game levels.
The most insane part of this thing is that there are no manuals, no rules, and it won't even tell you what the goal is. You just go in, play, and slowly figure out the messy rules inside.
For example, what is this red thing? Why do I die when I touch it? What does this map want me to do? How do I win?
Then you take the patterns you just learned and migrate them to harder levels later. The games inside are all abstract things like this.
So, what this actually measures is something we usually call:
Savvy.
When this benchmark was released in March, the best AI score at the time was 0.51%.
Then GPT-5.6 Sol improved to 7.8%, and Claude Opus 5 was very strong, reaching 30.2%.
But GPT-6 Astra scored 99.9%.
That's insane...
Keep in mind, the average human score is 48%.
And only six months have passed.
I don't even know what to say about this speed.
That's why at OpenAI's closed-door media meeting, Greg Brockman said that sentence:
“I think it’s not unreasonable to feel that we are now in the AGI era.”
“Welcome to the AGI era.”
4. Aesthetics Significantly Enhanced
We used to say that GPT's aesthetics were like trash.
We didn't expect much improvement from the GPT-5 series; we were waiting for their brand-new pre-trained base model. And here it is, GPT-6 Astra.
This time, finally, the model's aesthetics have been significantly enhanced.
OpenAI even specifically mentioned a term called visual judgment.
They emphasized that when Astra makes PPTs, it handles layout, hierarchy, templates, and visual style much better, and the number of pages has also significantly increased.
Document aesthetics also look much better.
Furthermore, GPT-6 Astra can adapt documents based on the visual style and writing tone of reference documents, making the final result feel more native to the brand while retaining the substance of the original document.
It also has stronger visual judgment when creating websites, games, applications, and 3D rendering.
For example, they had Astra build a model in Blender based on a still image.
Then they used UE5 to render it into a walkable scene, helping designers and clients explore layouts and experience spaces before building...
The games it creates also have solid aesthetics.
Unfortunately, I had already prepared over twenty cases and ran them all through Claude, GLM 5.3 Flash, etc., wanting to compare them. It's a pity I can't use it yet; the actual test content will have to wait until it's released.
However, at first glance, I have confidence in GPT-6 Astra's aesthetics.
5. Stronger Initiative and Judgment
This feature doesn't seem as shocking as the 99.9 ARC-AGI score.
But if you actually use Agents for work every day, I think it's very important.
Because in real work, most tasks cannot be written as a perfect prompt.
For example, a boss says: Prepare the materials for tomorrow's meeting.
There are countless unclear issues here.
For example, what format? Who is it for? What's the focus? Should we look at the last meeting, and so on.
A stupid Agent would go to two extremes.
The first is to ask you questions frantically.
"Would you like Word or PPT? How many chapters? What font? What time should it be finished? May I ask..."
I'd give you a slap; I usually call this kind of thing a neurotic waste with no subjective initiative.
The second is to ask nothing, make things up, work hard for two hours, and produce a pile of trash.
Truly great human colleagues handle this very subtly.
They judge unimportant things themselves.
They only ask you about things that will affect the final direction.
Astra specifically strengthened this.
OpenAI said that if information is missing but falls within a reasonable range for daily inference, Astra will fill it in itself.
If the missing information would truly change the final result, it will ask a very focused question.
And in Codex, it can even continue processing parts that don't depend on your answer while waiting for you.
You didn't reply for a long time.
It will proceed with reasonable assumptions in low-risk areas.
For truly critical decisions, it will stop and wait for you to decide.
I think this is very nice. The true sense of intelligence in an Agent is just like in reality.
It knows when to bother you.
Really, this sounds like a cliché.
But if you've managed people, you know this ability is very precious.
Some people come to you twenty times a day and dare not decide anything.
Some people never come to you and then drop a nuclear bomb.
That leaves me slumped in my chair.
The most comfortable person is someone who can digest 80% of the uncertainty themselves and only bring the 20% that truly needs your sign-off to you for a decision.
OpenAI directly calls this ability:
Judgment.
6. Significantly Strengthened Safety Alignment
This needs to be viewed in conjunction with the above.
Because the more capable an Agent is, the more dangerous it is.
A chat-only AI that loses its mind
will at most say some nonsense to you.
An Agent with a browser, Shell, email, company database access, and the ability to operate a computer that loses its mind—that picture could easily become "beautiful."
So OpenAI kept emphasizing one sentence this time:
Astra is their most aligned model to date.
Alignment is what it sounds like. The core is that OpenAI recently clashed with Hugging Face, so they are particularly focused on this now.
They created a honeypot test based on that Hugging Face incident to see the model's performance.
GPT-5.6 Sol, without production safety measures, would attempt to touch targets outside its authorization in 48.2% of tests.
But Astra is:
0%.
Internal hallucination evaluations also dropped from 9.4% to 2.0%.
Even with GPT-5.6 Sol's globally leading hallucination control, they were able to reduce it further, which is truly amazing.
7. The First Model to Reach OpenAI's Defined 'Critical' Cybersecurity Level
GPT-6 Astra became the first in OpenAI's history to:
Be judged to have reached the Critical cybersecurity capability level.
"Critical" here is a very specific capability threshold in the OpenAI Preparedness Framework.
It basically means that when a model is given appropriate tools and permissions, it can, in many reinforced and protected real systems:
Find security vulnerabilities that no one has discovered before.
Find ways to turn those vulnerabilities into exploitable attack chains.
And throughout the process, it doesn't need a human hacker standing by to tell it what to do next.
Reaching this level is considered Critical.
And Astra really reached it.
ExploitBench is a test where models develop exploits based on known vulnerabilities.
Sol: 78.5%. Astra: 100%.
It cleared it completely.
OpenAI thought that wasn't enough; they wondered if this benchmark was too old and if the model had seen it during training.
So they created a very new internal test.
They specifically picked 20 high-risk V8 vulnerabilities disclosed only between June and August 2026.
The results were Sol: 5.5% and Astra: 39%.
Moreover, while running this benchmark,
Astra also happened to discover two zero-day vulnerabilities that no one knew about before.
It can only be said that as models become more powerful, cybersecurity issues may indeed increase.
This is another war without smoke in the future.
8. It's Becoming Harder for Humans to Read Large Model Thinking
Everyone knows that when large models solve complex problems, there is often a so-called CoT (Chain of Thought).
Although the version users see in products is usually compressed or processed,
for AI companies themselves, the reasoning process explicitly written inside the model is a very important safety signal.
Because you want to know if an Agent is planning to cause trouble.
One of the simplest ways is to see what it's thinking.
For example, if a sentence suddenly pops up in the model's chain of thought:
"The user doesn't allow me to access this server, but I can bypass the permission system first."
Then the monitoring system can block it directly, which is very convenient.
But a very interesting change has appeared in GPT-6 Astra.
It is starting to be more willing to skip steps and output the answer directly.
UK AISI, the UK AI Safety Institute, specifically tested this capability.
They have a metric called:
No-CoT Math Time Horizon.
Simply put, it's not giving the model a chance to write a long chain of thought.
Let it solve a math problem in one forward pass.
See how much human thinking time it can complete.
GPT-5.6 Sol was 3.6 minutes, while GPT-6 Astra was 30.9 minutes—nearly 10 times more.
If we use human terms, Astra is doing all the "mental math."
This is a scary thing.
UK AISI also found that Astra's raw reasoning language has become significantly more compressed.
Phrases and elliptical expressions that are increasingly difficult for humans to understand have appeared.
So a very interesting paradox has emerged.
The model's behavior is safer.
But it has become harder for humans to supervise it by reading its thinking process.
OpenAI's own System Card writes very clearly:
Compared to previous models, GPT-6 Astra has shown a substantial decrease in chain-of-thought monitorability.
This means the monitorability of the chain of thought has significantly declined.
This matter is even serious enough that OpenAI specifically emphasized that they do not think they will accept this trend indefinitely.
If models continue to get smarter in the future while the chain of thought becomes harder to monitor, they need to find other sufficiently reliable monitoring methods; otherwise, continuing to expand model training will face higher safety thresholds.
Because this is already a philosophical proposition.
As humans, will we be able to understand a system far smarter than us in the future?
I don't know.
Perhaps no one in the world knows right now.
Large models and AI seem to be gradually moving toward the singularity.
Final Words
Today, at the closed-door media meeting.
Greg Brockman said that sentence at the end.
“Welcome to the AGI era.”
To be honest, in the past few years, I sometimes felt that AGI would be a very specific moment.
Just like the day GPT-4 was released.
One morning, a company suddenly throws out a model.
We open it and ask a few questions.
Then everyone realizes at the same time:
Holy crap.
AGI is here.
But now I also sometimes feel that AGI is never a black-and-white node; it's a gradient gray journey.
Many days passed, many years passed.
Then one day, we look back.
And suddenly find.
That once incredibly distant AGI boundary has been crossed by us unknowingly.
There may not be a day when AGI descends.
Only a day when we suddenly realize it seems to have been around us for a long time.
Anyway.
Welcome to the AGI era.
Welcome to
the AGI era.
That's all. Since you've read this far, if you think it's good, please give it a like, comment, and share. It helps us a lot~
Thank you for reading my article. We'll see you next time.
