GPT-6 Astra, Looped Transformers, and Hidden Reasoning
Ahead of AI Sebastian Raschka, PhD ● Covered by 25 sources
OpenAI’s GPT-6 Astra is out, and it’s very strong at coding, math, and computer use. The spicy part: reports say it may be using looped transformers and hiding its reasoning trace.
Based on reporting by Ahead of AI, Sebastian Raschka, PhD — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI’s GPT-6 Astra has landed with a lot of noise around it, and the model seems to have earned at least some of that attention. After using it for a couple of days, the source author calls it the best model they’ve used so far. It looks especially strong on 3D rendering, animation, and other graphical demos, while also jumping past GPT-5.6 in writing, math, coding, and more.
The benchmark picture is solid, but not magical. Astra is said to be strong on math and coding, and the headline number is its 99.9% score on ARC-AGI-3, versus 7.8% for GPT-5.6 Sol. Still, on independent evaluations like Artificial Analysis, it sits at the frontier rather than racing away from everything else. That matters because those benchmarks use different harnesses, and the author argues a model can look better or worse depending on the setup used to test it.
Computer use is where Astra really seems to stand out. The source describes it drawing in browser-based MS Paint with the mouse on the user’s computer, and treats that as more than a cute demo: it shows the model operating software through the Codex/ChatGPT app. The broader point is that models are moving beyond text and code into GUI work, which is useful for tasks that still live outside command lines and APIs.
There’s also a training angle here. The article says OpenAI reportedly bought tens of thousands of Mac Minis and Mac Studios, not to train on them directly, but to let models interact with macOS during reinforcement learning. The workflow is straightforward: give a task, show screenshots, predict clicks and key presses, execute them, show the updated screen, repeat, then use success or failure as feedback. NVIDIA’s CEO also said GPT-6 Astra was being trained on about 100,000 Grace Blackwell GPUs.
The architecture rumor is the more interesting wrinkle. The Information reported that Astra may use “recurrent depth” or “looped transformers,” which means reusing the same transformer blocks multiple times instead of adding new ones. That idea is not new; the article points back to Universal Transformers from 2018 and also walks through Nanbeige4.2-3B, where the same 22 blocks are passed through twice, giving 44 block applications without 44 separate sets of weights. The author’s bottom line is that Astra is still a reasoning model, trained with RLVR and producing intermediate reasoning traces, so the computer-use push does not amount to some new kind of model brain.
My take — AI-written commentary, not fact-checked reporting
The boring truth is that a lot of “secret architecture” talk is just old ideas wearing a new jacket. Looped blocks can be clever engineering, but they’re not wizardry, and hiding chain-of-thought is more likely a product choice than a scientific earthquake. The real story is simpler: models are getting better at doing actual computer work, which is far less sexy than a mystery and far more useful.
Read more about this at: Ahead of AI