Summary
Futurepedia tests GPT-6 Astra’s computer use feature across ten real-world applications, ranging from Google Earth Studio and Canva to Adobe Premiere, Photoshop, Illustrator, and Blender. All demos run through Codex inside the ChatGPT desktop app, which controls software by moving the mouse and typing — operating the interface the same way a human would — without scripting or APIs unless explicitly permitted.
Results vary significantly by application and task complexity. Google Earth Studio produced a high-quality custom animation after three follow-up prompts totaling roughly 35 minutes of compute time, demonstrating solid spatial reasoning across key frames and easing curves. Canva drawing tasks improved progressively over multiple iterations. In Illustrator, forcing Astra to use the visual interface (rather than the software’s built-in Python API) took over 90 minutes and produced a poor result; allowing the Python path achieved a better outcome in under 30 minutes. This comparison runs throughout the video: Astra will default to programmatic shortcuts in Blender and Illustrator when not restricted, and those paths generally outperform pure computer use for complex creative tasks.
The video offers useful calibration for builders and power users evaluating where GPT-6 Astra’s computer use genuinely saves time. The practical takeaway is that the feature works well for moderately scoped tasks in interfaces with clear visual affordances, and pairs best with a human doing final refinement rather than running unsupervised to a finished result.
📺 Source: Futurepedia · Published September 15, 2026
🏷️ Format: Review
