When OpenAI’s GPT-6 Astra started giving noticeably shorter chain of thought, the visible step-by-step reasoning a model prints before its final answer, a report claimed the model was deliberately hiding that reasoning from users. AI researcher Sebastian Raschka says that explanation gets the architecture wrong.
What he said
Raschka calls the hidden-reasoning theory “confused reporting.” His alternative explanation: Astra likely reuses a technique called a looped transformer, where the model runs the same set of internal layers more than once on a problem instead of building a longer chain of separate layers.
He points to a similar architecture in the open-source model Nanbeige as a working example, where “the hidden states are fed back through the same 22 blocks” a second time, doubling the effective depth of the reasoning without doubling the number of trained layers. Shorter visible reasoning, in his view, doesn’t mean the model is concealing steps. It can just mean the model needs fewer of them: “Using fewer tokens could just mean that the model is more capable and makes fewer mistakes, uses less backtracking.”
Who he is
Sebastian Raschka, PhD, is an AI researcher and author of books on building and understanding large language models, including “Build a Large Language Model From Scratch.” He writes the Ahead of AI newsletter, where he regularly breaks down new model architectures in technical detail.
What he gets right, and where it’s incomplete
Raschka’s technical distinction holds up. A model that loops through the same layers twice is doing extra computation, not hiding a reasoning trace. That’s a real and long-studied architecture choice, not a cover story. He also isn’t alone in pushing back on the concealment framing: OpenAI’s chief scientist has said chain of thought monitoring remains a priority internally, and that any drop in how easily a model’s reasoning can be tracked stems from factors unrelated to this kind of architecture change.
Where the piece is thinner is on certainty. Raschka is inferring Astra’s architecture by analogy to Nanbeige and other published looped-transformer research. OpenAI hasn’t confirmed Astra actually works this way, so this remains an informed hypothesis, not a confirmed technical fact about a closed model.
Why it’s notable
BuilderWithin covered Astra’s capabilities and Copilot pricing six days ago when it launched. This is a different question: not what Astra can do, but whether builders should trust what it shows them while it works. That question matters more as coding agents make more autonomous decisions during long tasks, and users increasingly rely on visible reasoning to catch a mistake before it ships.
What it means for builders
Don’t treat a shorter reasoning trace as evidence a model is hiding something from you. A capable model doing a task efficiently and a model concealing its steps look identical from the outside, which is exactly why Raschka’s architectural explanation matters: it gives you a specific, checkable reason for the behavior instead of a suspicion.
The practical takeaway is to keep judging a model like Astra by its output quality and by whether it still lets you inspect its work when something goes wrong, not by the raw length of the reasoning it shows you.
End of article