GPT-6 Astra: The Hidden Chain of Thought Hides No More Than What the Architecture Lacks

Perex: The GPT-6 Astra model brought better results, but also concern about shorter and hidden chains of thought. However, a detailed look at its probable architecture explains why the model can do without a lengthy internal dialogue.

A looped transformer recycles blocks instead of adding new ones

According to technical speculation discussed by researcher Sebastian Raschka, GPT-6 Astra utilizes the concept of looped transformers. Instead of processing data in a long row of single-use blocks, the model sends intermediate calculations back through the same set of blocks. The weights in these blocks remain the same, giving the model a so-called recurrent depth without a dramatic increase in the number of parameters. It's not a new idea (it already appeared in Universal Transformers in 2018), but it seems OpenAI has found a way to streamline it. Within the same compute budget, the model can achieve better results than a traditional architecture thanks to looping.

Less reasoning is not necessarily withholding evidence

One of the main criticisms against GPT-6 Astra is reduced monitorability. The model generates shorter chains of thought than its predecessors. But Raschka argues this doesn't necessarily mean strategic withholding. More capable models simply make fewer mistakes and don't need to try as many dead ends. A shorter Chain of Thought may simply be a consequence of the model finding the right solution earlier. After all, the same trend already applied to earlier models. The smaller Luna required more reasoning tokens than the more capable Sol to achieve a similar result. Looping adds computational power inside the architecture itself, reducing the need to use external reasoning tokens as an improvised scratchpad.

The computer as an environment, not just hardware

While the discussion revolves around architecture, Astra's true breakthrough is its ability to control a graphical user interface (computer use). Raschka points out that OpenAI's massive Mac purchases weren't for training directly, but to create an interactive macOS environment. The model learns to recognize the screen and generate mouse or keyboard actions through reinforcement learning. This makes the model a much more versatile tool for tasks that lack their own API interface.

The push for different tools will reveal the limits of current practices

Expectations for future models should shift from abstract reasoning to interaction with real systems. For developers, this means one thing. They will have to throw out old instructions (prompts) and delete the lengthy manuals (SKILL.md) that guided older models step by step. New models no longer need this help, and too detailed instructions would only limit them. The question remains how quickly open-source solutions can mimic not only the architecture but also the training infrastructure necessary for smooth UI control.

Lilith's verdict

If shorter reasoning comes from internal network loops, Astra is not necessarily sneakier, it’s just better at doing mental math rather than writing it out. But in both cases, you get a much narrower window into the model’s head.