GPT-6 Astra, Looped Transformers, and Hidden Reasoning

Ahead of AI published an article examining OpenAI's GPT-6 Astra, which the article states was released last week, along with reported looped transformer or recurrent depth architecture rumors and speculation that Astra hides its reasoning trace, or chain of thought.

The author writes that Astra is exceptionally good, likely the best model used as of writing, and disproportionately strong at 3D rendering and animation tasks relative to other models, leapfrogging its GPT-5.6 predecessor in writing, math, coding and graphical demos. The article states Astra achieves 99.9% on the ARC-AGI-3 benchmark, which measures logic puzzle solving and generalization, versus 7.8% for GPT-5.6 Sol. In the Artificial Analysis Coding Agent Index v1.4 and the Artificial Analysis Intelligence Index, the article says Astra is at the frontier but does not pull ahead by leaps and bounds. The article notes Artificial Analysis benchmarks are independent and may be more trustworthy than self-evaluated benchmarks, and that harness choice can affect agentic evaluation results.

The article describes Astra as strong in image and rendering tasks and in computer use, operating software on a local computer through the Codex/ChatGPT app. It also references reporting that OpenAI purchased tens of thousands of Mac Minis and Mac Studios for reinforcement learning, with macOS serving as a training environment rather than the training hardware, and mentions that NVIDIA's CEO said GPT-6 Astra was trained on roughly 100,000 Grace Blackwell GPUs. The article states Astra remains a reasoning model trained with reinforcement learning with verifiable rewards that produces intermediate reasoning traces.

The article also explains looped transformers, citing the 2018 Universal Transformers paper and the Nanbeige4.2-3B open-weight model released in July, which applies 22 transformer blocks twice using shared weights. It also cites ByteDance's Ouro-Thinking 2.6B, which applies a 48-block stack four times, and the adaptive halting idea from Universal Transformers. It reports that The Information published an article about two days before the official model stating, according to inside information, that Astra uses recurrent depth or looped transformers.