OpenAI's GPT-6 Astra arrived last week to predictable fanfare, but the substantive story sits beneath the demo clips. The model scores 99.9% on ARC-AGI-3 (up from GPT-5.6 Sol's 7.8%), leads on math and coding benchmarks, and dominates 3D rendering and animation tasks in a way no predecessor has. On the independent Artificial Analysis Intelligence Index and Coding Agent Index v1.4, Astra sits at the frontier — though notably without the dramatic separation from competitors that the self-reported numbers suggest. The gap between self-evaluated and independently-evaluated performance remains a persistent credibility tax across the industry. The architectural headline is the rumored adoption of "recurrent depth" or looped transformers. The core idea is straightforward: instead of stacking hundreds of unique transformer blocks, the model passes intermediate representations through the same set of blocks multiple times. This trades parameter count for adaptive compute depth — the model can "think longer" on harder inputs by looping more, without proportionally growing the weight count. It is not a new idea in the literature, but deploying it at GPT-6 scale would be a genuine engineering milestone, and The Information's reporting suggests OpenAI has done exactly that. The computer-use training pipeline is arguably more consequential than the architecture shift. OpenAI purchased tens of thousands of Mac Minis and Mac Studios — not for training compute (that runs on an estimated 100,000 Grace Blackwell GPUs, per NVIDIA's CEO), but as RL environments. The training loop is clean: prompt the model with a GUI task, feed it screenshots, let it predict mouse and keyboard actions, execute those actions on the Mac, capture the new screen state, and use success/failure signals as RL reward. This is standard RLVR (Reinforcement Learning with Verifiable Rewards) adapted to a visual-spatial action space. The investment signals that OpenAI views GUI-native operation — not just API and CLI fluency — as a core product capability, not a demo gimmick. The "hidden reasoning" rumor deserves careful handling. Astra is still a reasoning model producing intermediate chains of thought, trained via RLVR. The looped transformer architecture does not inherently hide reasoning traces — loops are an inference-time compute mechanism, not a chain-of-thought suppression mechanism. The author of the original article (who specializes in LLM architectures) explicitly flags this conflation. Whether OpenAI is separately choosing not to expose reasoning tokens to users is a product decision, not an architectural consequence of looped transformers. The independent benchmark picture is more nuanced than the headlines. Artificial Analysis uses shared harnesses (like Stirrup) across models, which enables apples-to-apples comparison but introduces a systematic bias: models are typically fine-tuned for their primary harness, so third-party harnesses may understate performance. The Intelligence Index v4.2 blends Terminal-Bench v2.1, τ³-Banking (τ-Bench harness), and other evaluations. Astra leads, but the margin over competitors in this blended view is thinner than in OpenAI's own reporting. A practical footnote worth flagging: as models improve at task comprehension, the accumulated AGENTS.md and SKILL.md instruction files that developers have built up may now constrain rather than help newer models. The recommendation from multiple sources (including Claude Code's lead) is to archive old instruction files and let the model generate fresh approaches. This is a small but telling signal — the models are outgrowing the scaffolding their users built for weaker predecessors. The 20-year trajectory here is clear: the convergence of looped-transformer architectures (adaptive compute depth), GUI-native RL training (Mac farms as environments), and reasoning-model post-training (RLVR) points toward models that operate computers the way humans do — visually, interactively, across arbitrary software. The question is who captures the value. If computer-use capability remains locked inside proprietary harnesses (Codex, ChatGPT app), the generative potential is enormous but the access bottleneck is OpenAI's product layer.
