GPT-6 Astra: strong on computer use and graphics, trained on a wall of Mac Minis

Sebastian Raschka’s early hands-on take frames OpenAI’s newly released GPT-6 Astra as the strongest model he has used, with an outsized lead in graphics-heavy work like 3D rendering and animation. It tops predecessors across writing, math, and coding, and posts eye-catching benchmark numbers — including 99.9% on ARC-AGI-3 versus 7.8% for the prior GPT-5.6 generation. On independent, harness-normalized suites such as the Artificial Analysis Intelligence and Coding Agent indices, however, Astra sits at the frontier without running away from the pack. Raschka cautions that agentic scores are sensitive to the test harness, and because models are typically tuned against one primary harness, cross-harness comparisons may understate a given model’s real-world strength.

The more consequential shift is computer use: Astra operates local desktop software directly through the Codex/ChatGPT app, driving GUIs with the mouse and keyboard rather than relying on CLIs or APIs. Raschka connects this to reports that OpenAI bought tens of thousands of Mac Minis and Mac Studios — not to train the model (GPUs still do that) but to expose real macOS as a reinforcement-learning environment. The training loop is a screenshot-action cycle: the model is given a task, sees the screen, predicts clicks and keystrokes, the harness executes them on the Mac, and success/failure verifiers supply the reward signal. He argues GUI control is the machine equivalent of a humanoid robot — less efficient than a purpose-built interface, but general enough to cover the vast amount of software that never exposes a command line.

The piece also teases its technical core — looped transformers and recurrent-depth architectures — and the rumor that Astra hides its chain-of-thought reasoning trace, which Raschka promises to examine alongside recent research. A practical aside worth flagging for engineering teams: he echoes advice attributed to the Claude Code lead that stale AGENTS.md and SKILL.md instruction files may now hurt more than help, since newer models can over-fit to hand-holding scaffolding and would often find better solutions if left to reason freely.