GPT-6 Astra: what OpenAI’s new model can actually do

Robbert van Empel has published an eight-minute video about GPT-6 Astra, focusing on five examples that make the model’s capabilities easier to see than benchmark tables alone. The video went live on September 4, 2026, one day after OpenAI announced the model. Van Empel is transparent about the scope: he did not yet have access to Astra when recording, so this is a guided review of published demonstrations, not an independent hands-on test.

The video first places Astra’s results in context. OpenAI reports a 99.9% score on ARC-AGI-3 and substantial gains in computer use and professional work. These are company-reported evaluations and depend on the test setup. The video therefore moves quickly to visible outputs: a playable 3D mini-planet reportedly generated in one shot, an interactive environment that lets visitors walk through a Van Gogh-inspired town, and a Blender scene used to compare Astra with Anthropic’s Claude Fable 5.1.

Two larger demonstrations show why the release may matter to creators and developers. One presents a Manhattan environment built in Unreal Engine, complete with streets, traffic and moving pedestrians. Another places autonomous AI agents in a simulation where they begin developing ways to communicate. The examples suggest that frontier models are getting better at combining code, visual judgment and multiple steps into artifacts people can inspect or use. They do not show a general success rate, how much human correction was needed, or whether other users can reproduce the same results.

That distinction is important. A polished demo can reveal what is possible under selected conditions, but it is not the same as a controlled comparison. Prompts, run time, tools, hardware and editing all influence the outcome. The video’s practical value lies in bringing several early examples together and showing them on screen, while the linked sources in the description let viewers explore the projects and original posts for themselves.

For AI users, makers and educators, the clearest takeaway is that Astra is aimed at finished digital work, not only text responses. Games, 3D scenes and agent simulations can become useful teaching and prototyping material when their limitations remain visible. Teams considering Astra should test their own representative tasks, document the human interventions and keep review points before generated work is published or deployed.