Summary
GPT-6 Astra, OpenAI’s latest flagship accessible via the ChatGPT desktop app and the platform formerly known as Codex, gets put through a series of demanding creative and technical challenges in this video from AI Search. The tests are deliberately varied and difficult, designed to expose the model’s reasoning, physics understanding, and multi-step execution rather than cherry-picked easy wins.
The opening challenge asks Astra to build a ray-traced simulation of a bullet piercing a water balloon — fully from scratch, no external libraries — with adjustable sliders for bullet speed, balloon size, gravity, surface tension, and spray density, viewable from any camera angle. A self-critiquing agent loop is embedded in the prompt: a critic agent takes screenshots each round and scores realism from 0 to 10, with the builder iterating until it hits 8 or above. After 32 minutes and three rounds (scores: 4.7, 6.0, 6.5), the result is a convincing interactive simulation. Subsequent tests include a fully playable first-person open-world RPG built without asset libraries, a Euro EDM track composed and mixed in a Waveform DAW, and — most impressively — a live classical piano performance where Astra presses keys sequentially with correct timing inside a DAW, producing a result indistinguishable from a human performance in roughly seven minutes.
Throughout, the video demonstrates a practical prompting pattern: embedding a critic agent into any generative loop to force iterative improvement rather than accepting first-pass output. For developers and power users evaluating what GPT-6 Astra can realistically accomplish on complex, open-ended tasks, this is one of the more technically imaginative assessments available.
📺 Source: AI Search · Published September 07, 2026
🏷️ Format: Showcase
