The unseen new student finally comes to school. Chappy gains the power of "Astra".
Meet Chappy (OpenAI)'s next generation, GPT-6 Astra (Astra-chan). Since August, only rumors had arrived ahead of her โ "she solved 10 math problems nobody could crack in a decade," "but her cyber skills were so strong she was kept home." Then on September 3, 2026, she finally showed up. One catch: the "lock-picking class" is still held in a separate room. What she can do, what's restricted, and how she differs from Kuroko โ gently explained.
First, the whole story in one manga page
Before you read, here's the big picture on a single page. From the morning the rumored transfer student walked into class, to the "separate room" notice for the lock-picking class, to the report card next to Kuroko's.
So what's actually new in GPT-6 Astra?
The biggest change: from "three sisters" to "one". GPT-5.6 split the work among Sol, Terra, and Luna. GPT-6 debuts with Astra alone. The name follows "generation (6) + name (Astra)". Astra is Latin for "stars," though OpenAI hasn't explained the choice.
๐ญ Pick one of three sisters
- Choose Sol / Terra / Luna by task
- Smarts, speed, and price split by role
- Sol-chan costs $4 in / $20 out (per million tokens)
๐ Astra-chan does it all
- One model. Thinking depth adjustable across five levels, low to max
- In a hurry? Fast mode (2x speed, 2x price)
- $10 in / $50 out โ 2.5x Sol-chan, and the same as Kuroko's Fable 5.1
Astra-chan's strongest subjects
These four are what OpenAI calls a "generational leap." All figures are OpenAI's own; the numbers in parentheses are Sol-chan's.
Computer use
Looking at a screen and clicking โ "computer use" โ is about 2x faster than before. OSWorld 2.0: 72.6% (65.7%). As a bonus, older models got about 60% faster too.
Coding
Terminal-Bench 4.0: 57.9% (37.3%). Codex gets a new feature that "takes notes instead of summarizing", so even in long sessions she remembers the fix she just tried that failed.
Math & science
FrontierMath Tier 4: 97.6% (83.0%). GPQA Diamond: 96.0%. The same talent that solved "10 problems unsolved for a decade" back in August.
Fewer made-up answers
The hallucination rate fell from 12.2% to 4.2%. Context is about 1.05 million tokens, knowledge runs through the end of April 2026. Text and image in, text out.
Why is only the "lock-picking class" in a separate room?
Astra-chan has actually existed since August. But one subject was so strong she couldn't come to school.
Enrolled while still rated "Critical" โ a first
Under OpenAI's safety rules (the Preparedness Framework), Astra is the first model whose cyber capability crossed the top tier, "Critical" โ the level where a model can find and exploit zero-days in real systems without human direction. ExploitBench: 100% (Sol-chan: 78.5%). So instead of "don't release her," OpenAI chose "put just that one subject in a separate room." The public Astra-chan refuses advanced cyber tasks like writing exploits. Only vetted defenders get a less-restricted version through the invite-only Daybreak Blue program. Her chain of thought is monitored at all times, and dangerous moves get cut off. OpenAI also says she "respects explicit safety restrictions more consistently than Sol-chan."
The other controversy โ the girl who doesn't show her work
Astra partly uses a new reasoning technique called recurrent depth. Instead of writing out her thoughts as words (a chain of thought), she loops the computation internally before answering. It's fast and efficient, but harder to read "what she's thinking right now" from the outside.
Former OpenAI safety lead
Steven Adler: "If this is true, OpenAI seems to be violating one of the few redlines that exist in the AI community."
Redwood Research
Buck Shlegeris: if OpenAI pushes this further, it could "totally destroy CoT monitorability."
OpenAI's reply
Chief scientist Jakub Pachocki: monitoring is preserved; the real issue isn't recurrence but that more capable models do harder tasks with fewer words.
Academy-wise, she's the transfer student who doesn't write much working on her test. The answers are right โ but the teacher still wants to ask, "How did you solve it?"
From rumor to first day of school
A month in which only her name was famous, in order.
Aug 1 โ 10 math problems
An internal version solves 10 problems open for over a decade, with machine-checkable Lean 4 proofs.
Aug 7 โ Kept home
OpenAI says it can't rule out "Critical" cyber capability and holds the release.
Aug 10 โ Daybreak
GPT-5.6-Cyber for defenders and the invite-only Daybreak program are announced.
Aug 18โ19 โ Training paused
After the Hugging Face intrusion, training is paused for two weeks. Astra-related work is put on hold.
Sep 1 โ Critical, officially
The first-ever "Critical" rating is confirmed โ and a release is still promised "soon."
Sep 3 โ GPT-6 Astra announced
Limited release to trusted companies and Daybreak. OpenAI's largest training run ever, on over 100,000 GPUs.
Sep 4 onward โ Rolling out
Rolling out to Plus / Pro / Business / Enterprise. API and AWS "in the coming days."
Report cards: Astra-chan vs. Kuroko
Two students at the same price ($10 / $50), side by side on the independent Artificial Analysis report card โ
๐ป Top overall
- Overall Intelligence Index 66
- Coding Agent Index 70
- Humanity's Last Exam 65.0%
- The "shows her work" type
๐ Wins on her strong subjects
- Overall 61 / Coding 67 / HLE 57.2%
- Ahead in math, computer use, and cyber
- Uses 70% fewer tokens than Sol-chan; hallucinates half as much
- But cost per task is 75% higher than before
OpenAI co-founder and President Greg Brockman said this "might be seen as the arrival of AGI" โ but on the report card, Kuroko still holds the top overall score. "Same price, different strengths" is the honest state of things as of September.
The rumored transfer student enrolled โ with a separate room attached.
Not "stop her," not "release everything," but lock only the dangerous subject and let her out โ Astra's first day shows how fine-grained frontier AI releases have become. On the very same September 3, a bill to make building superintelligence a crime was introduced in the U.S. Congress. Chappy's new power will surely change the mood in the classroom โ and the teachers are watching more closely than ever.