GPT-6 Astra workflow reportedly turns 20 property photos into a walkable Unreal Engine 5.8 scene

• A reported 21-second prototype turns 20 photos from an OLX property listing into an interactive Unreal Engine 5.8 interior, according to the original walkthrough.
• The working route moved into a locally installed Unreal build connected to Codex after Blender tests fell short, AmirMushich's Blender reply says.
• The project remains a demonstration, not a released kit: AmirMushich's workflow reply says he is considering a prompt and workflow explanation as a reusable artifact.
• The run consumed 259.6 million tokens, nearly all cached, with an estimated API cost of about $398, per AmirMushich's token tally.

The exterior was not reconstructed from the interior photo set. The original walkthrough says Google Maps panoramas at relevant coordinates supplied the surrounding context, while AmirMushich's frame-rate reply puts the live run at 150fps and the posted render at 50fps.

Mushich described a three-part input and output chain:

1. Interior reference: 20 photos from the OLX listing.
2. Exterior context: Google Maps panoramas tied to relevant coordinates.
3. Experience: a first-person, interactive scene in Unreal Engine 5.8.

The resulting clip cuts from a room photograph to a navigable version of the space. The source material was limited to interior photos, with map panoramas filling window views and surroundings, AmirMushich's reply specifies.

Mushich says Blender was tested first, then abandoned for this project. He installed Unreal locally and connected Codex to the engine, turning the model into an operator inside the target scene-building environment.

That setup matches OpenAI's Astra overview, which describes the model working across desktop applications and directing 3D work in Blender. Its Codex prompt guidance also argues for trimming bloated agent instructions rather than piling on more scaffolding.

Mushich expects the differentiator to remain the people who master these tools, as he put it in a reply.

The published walkthrough is not positioned as an accurate property survey. Mushich said the prototype was not 100% accurate and that a real listing would need property documents and ideally an interior designer to check errors, in a reply.

The visible performance figure is also a split between the interactive build and the export: 150fps in the original run, 50fps in the rendered video. Mushich had tested immersive VR with Unity and Android phones around 2018, his VR note says, but this is the first public demonstration he has described with Astra and Unreal.

Mushich's project breakdown lists:

• Total: 259.6M tokens
• Cached: 248.9M tokens
• Output: 1.07M tokens
• Estimated current API cost: about $398

He said the work fit within a $200-per-month subscription, without publishing the task log or how that plan usage mapped to the API estimate. OpenAI's usage guide says Work and Codex share a plan allowance whose coverage varies with the plan, task, model, and settings.

Big agentic environment work can also burn a budget without producing a usable scene. AIandDesign's account describes spending 40% of a $200 Codex plan and substantial Muse usage on a dynamic-environment experiment that never landed.

The public artifact is still the video, not a step-by-step project. Mushich said he was normalizing the workflow before sharing it in another reply, and floated a prompt plus workflow explanation rather than a finished tool.

AmirMushich gave a one-word answer in a reply, while gregisenberg posted a link-only share; neither offered implementation details. The evidence identifies the inputs, the Unreal and Codex handoff, the performance figure, and the reported token spend, but not the automation scripts or asset-generation steps between them.