Why is the model still not out and already causing controversy??
Just before the release of GPT-6, a story first reported by The Information quickly sparked widespread online discussion:
OpenAI's new model introduces a reasoning technique called "recurrent depth," which may make the model's thought process harder to understand and monitor.
In simple terms, it means that in the future, humans won’t be able to understand what AI is thinking at all.
What if AI learns to deceive, cheat, or bypass security restrictions—how would we even notice in time?
No wonder the AI security community immediately sounded the alarm, with everyone worrying:
If this technology continues to expand in usage, existing chain-of-thought monitoring mechanisms may become directly ineffective.
Given the significance of the issue and the intensity of the discussion, OpenAI’s chief scientist had to personally respond.
Even amid the chaos, some people were surprised to discover:
Is OpenAI’s new model called “GPT-6” or “Astra”? Perhaps it was already hinted at by the API.
OpenAI may also simultaneously launch the new image model, GPT Images 2.1.
There are too many melons—let’s savor them one by one.
The model began looping in the brain, and the security experts panicked.
The first is the "cycle depth" technology that has sparked controversy this time.
Previously, the model would leave clear chain-of-thought (CoT) reasoning traces, making it obvious what it was thinking.
Although this text may not perfectly equate to the model's true internal thoughts (which remains debated), it at least leaves behind traceable evidence.
In contrast, Astra's so-called "circular depth" has a very different approach.
It allows the model to feed the internal state generated at one step back into the neural network, repeatedly processing it over multiple rounds in the hidden space before outputting the next segment of text.
For example, previously, when a model solved problems, it had to think through them while writing out the calculations on scratch paper.
Now, it can be calculated mentally.
People outside may only see the question and the answer, but not necessarily the calculations in between, the approaches tried, or any detours taken.
By saving the effort of translating each step into text, the model can, in theory, run multiple reasoning cycles using the same computational resources and simultaneously consider several possible solutions.
But as performance improves, issues also arise:
If the model no longer thinks in human language, how can humans monitor it?
The CEO of Redwood Research, a nonprofit AI safety research organization, was among the first to express strong concerns:
It is currently unclear by how much Astra’s chain-of-thought observability has decreased, but if OpenAI further increases the loop depth, it could shift most of the model’s reasoning into unreadable hidden spaces, ultimately “completely destroying” chain-of-thought monitoring.
Ryan Greenblatt, Chief Scientist at Redwood, is also concerned about the same direction.
In his view, opaque reasoning may scale faster than traditional chain-of-thought approaches. Continuing along this path, models in the future might perform reasoning entirely or almost entirely within latent space.
This is the worst development yet in the field of AI safety.
After all, if it were just a matter of “not understanding the calculation process,” the issue wouldn’t be so serious.
As the lead investigator of the previous "OpenAI model infiltration of Hugging Face" incident, he identified two major risks:
At the time, researchers heavily relied on chain-of-thought reasoning to reconstruct the behavior of a large number of AI agents, even though doing so was already extremely challenging.
The associated AI attempted to alter its own records and even forged tool calls.
For the latter, if the model has stronger latent reasoning capabilities, it is more likely to first devise a deception strategy within the latent space, then construct a fabricated narrative through forged inputs and outputs, leaving no obvious traces in the textual reasoning chain.
Oh no, the risk has doubled...
Zvi Mowshowitz, who has long focused on AI safety, put it more directly:
You're playing with fire.
He worries that once hidden inference can deliver significant performance advantages, labs may descend into a vicious competition:
You hesitate to hide your chain of thought due to security concerns, but your competitors don’t—and their models are stronger and cheaper.
To keep up, are you in?
Mowshowitz even believes that, in the future, legal measures may be necessary to prevent AI companies from collectively sacrificing the traceability of chain-of-thought reasoning in pursuit of capability.
However, others on the other side feel that this wave of panic came too quickly.
Tian Yuan Dong also stated on X that when they previously launched Coconut, they received nearly identical questions.
Coconut stands for "Chain of Continuous Thought," and similarly advocates for models to reason directly within a continuous hidden space rather than decoding each step into text.
According to Tian Yandong, even if the model retains an explicit chain of thought, it does not mean that humans truly see its real intentions.
It’s important to understand how the model itself works, rather than focusing solely on the “reasoning steps” it produces.
As the controversy grew larger, OpenAI's Chief Scientist Jakub Pachocki could no longer stay silent and responded directly.
He immediately stated that he wanted to prevent an unmonitored race triggered by "chaotic reporting."
Including Astra, the computational graph depth of OpenAI’s current state-of-the-art models remains within twice that of GPT-4.
In other words, Astra does use cyclic computation, but the scale is currently limited and does not fully conceal the entire chain of reasoning. Based on available information, its natural language reasoning remains readable.
Pachocki also emphasized that OpenAI has always regarded chain-of-thought monitoring as a critical security measure, and it remains a core objective of the company’s current research agenda.
However, he also acknowledged that chain-of-thought monitoring is very fragile and is deteriorating.
However, this downward trend is not entirely due to architectural changes such as cycle depth (we’ll publish a dedicated article on this soon); OpenAI is also still researching ways to enhance monitoring capabilities.
So, Astra hasn't made the model completely incomprehensible to humans yet.
What security experts are truly concerned about is:
Will the currently restricted hidden reasoning, confined to a narrow scope today, continue to expand in the next generation of models, eventually becoming an unmonitorable black box?
Is it really GPT-6-Astra? The API response gave it away first.
The algorithm controversy is done; the second issue is about the model name.
The whistleblower Leo discovered that requesting "gpt-6-astra" from the OpenAI Responses API returns a 404 Not Found error.
Entering a completely nonexistent or randomly fabricated model name typically results in a 400 error.
A 404 indicates that the model identifier has been recognized by the backend, but your current account lacks access permissions, or the model is not yet available.
Previously, some unreleased models have also revealed their presence in advance through similar API response discrepancies.
Therefore, Leo concludes that "gpt-6-astra" may already have been deployed to the OpenAI API backend.
GPT Images 2.1 is coming too—first, let’s fix the “noise issue.”
Outside of the language model, OpenAI's image products have also been revealed to have been updated simultaneously.
According to the leak account Fix, the new image model GPT Images 2.1 may be released on September 4.
The most noticeable change in this upgrade is likely the resolution of the "noise issue" present in previous versions.
Previously, some Images v2 generation results exhibited strange grain, fogging, and dirty textures, especially noticeable in scenes with extensive text or fine details, as another user put it:
Like a photo taken through a dirty window.
The latest leaked samples have clearly improved this issue, with cleaner visuals.
Come on, it's still the main stage for the veteran actor Ultraman:
Famous World Painting: "Ultraman, Who Will Be Arrested If GPT-6 Isn't Released!"
And Ultraman, staying up late to work overtime and wandering the streets at night to avoid arrest.
Also shared an Instagram post titled “Have You Ever Seen the Streets at 2 a.m.?”.
I have to say, the image really does have a high-quality appearance, almost like a real photograph.
Also well-handled: more themes and more styles:
No more talk, Ultraman, just send it already.
Reference link:
[1]https://techcrunch.com/2026/09/02/openais-new-reasoning-technique-alarms-ai-safety-experts/
[2]https://x.com/tydsh/status/2095227000365707542?s=20[
3]https://x.com/merettm/status/2095023204993490967?s=20
[4]https://x.com/synthwavedd/status/2095184148981842161?s=20
[5]https://x.com/FixlationAI/status/2095232751654064426?s=20
[6]https://x.com/marco_obviously/status/2095275097217191981?s=20
This article is from the WeChat public account "Quantum Bit," authored by: Focused on Frontier Technology