OpenAI’s GPT-6 Astra is out and widely called the best model yet, but the architecture rumor overshadowed the benchmarks: reports said it runs on a “looped transformer,” with some suggesting the design exists partly to hide the model’s step-by-step reasoning. Sebastian Raschka — ML researcher and author of the Ahead of AI newsletter — wrote the clearest explainer yet of what looped transformers actually are. His verdict: the technique is real, probably in use, and mostly benign — but it sits inside a genuinely worrying trend about how legible frontier models remain.

His hands-on impressions:

  • Astra is the strongest model he has used, with its biggest gains in graphics and “computer use” — driving software with a mouse and keyboard, not just generating text. Independent coding-agent benchmarks put it at the frontier, but only modestly ahead of rivals.
  • One standout number: roughly 99.9% on ARC-AGI-3, a test of solving novel logic puzzles, versus 7.8% for its predecessor GPT-5.6 Sol.
  • Behind the scenes, OpenAI reportedly bought tens of thousands of Mac minis — not as GPUs, but so the model could practice using a real macOS desktop during training.

What is a looped transformer? A normal large language model is a stack of layers, and each piece of text passes through every layer once. A looped transformer reuses the same layers: data cycles through them several times with the same weights, buying extra computation — more “thinking” — without storing extra parameters.

  • The idea dates back to the 2018 Universal Transformer; recent open-weight models run whole stacks twice (Nanbeige) or four times (ByteDance’s Ouro).
  • Newer variants make the loop count a routing decision: Mixture-of-Recursions learns to send harder tokens through one, two, or three passes.
  • The research verdict so far: at matched compute budgets, looping gives a modest quality edge at scale — one September 2026 study estimates 6.8–18% less training compute to reach the same loss. It adds compute, not memory: memorization capacity stays flat with fixed parameters while multi-step reasoning improves.

Does looping hide reasoning? Chain-of-thought — the visible step-by-step working — has been hidden from users since OpenAI’s o1, so the real stakes are internal: safety teams read traces to check a model is behaving. The rumor is that looping moves thinking inside the architecture, leaving shorter, less useful traces. Raschka is skeptical:

  • At equal accuracy, Astra emits fewer reasoning tokens than its predecessor — but the same pattern held between the smaller and larger GPT-5.6 models, where the bigger one used about 80% fewer tokens at similar performance. More capable models make fewer mistakes; shorter traces track competence, not concealment.
  • A reasoning trace was never a faithful record of internal computation anyway. Astra’s system card does note reduced trace monitorability — mostly shorter, less informative traces — but nothing ties that to looping.
  • OpenAI’s chief scientist called the looping story “confused reporting,” while conceding that chain-of-thought monitoring is fragile and “trending in a negative direction” for reasons he says are unrelated to architecture.

The essay is a good template for reading frontier-model scoops: separate the plausible technical change (recurrent depth is real, increasingly common, and pays a modest efficiency dividend) from the alarming story bolted onto it (secret reasoning obfuscation). The genuinely disquieting signal is not the looped transformer at all — it is OpenAI’s own admission that its primary interpretability tool is quietly degrading.