Summary
Wes Roth shares an early hands-on demonstration of OpenAI’s GPT-6 Astra after roughly 12 hours of access, much of it running autonomously overnight. The central showcase is a complete 3D video game development pipeline orchestrated entirely by Astra: generating concept art with GPT Image 2.0, building and texturing 3D assets inside Blender, adding walking and movement animations, importing into Unreal Engine or Unity, integrating voices and sound effects via the ElevenLabs API, and finally play-testing the finished game autonomously in a live window — navigating corridors, interacting with NPCs, and verifying quest progression.
Roth walks through three distinct environments Astra constructed to specification: a post-apocalyptic wasteland with a hand-built rusted car, an alien comic-book world with ink-outline aesthetics, and an Escape from Tarkov-inspired urban ruin with Russian-style license plates and battle-damaged vehicles. He also demonstrates Astra ordering groceries from Instacart, trimming silence gaps in a video file using an actual video editor, building a playable RimWorld clone from scratch in Unity after analyzing open-source references, and creating a harness that allows Astra to play the commercial RimWorld title.
A key behavioral observation throughout is Astra’s persistence: the model pauses and waits through a roughly 20-minute GitHub asset upload rather than timing out, manages its own Codex token budget by drawing on pre-authorized reset pools, and operates with less back-and-forth than prior models. Roth frames the implications as significant for game development studios and any multi-tool workflow requiring long-horizon autonomous execution.
📺 Source: Wes Roth · Published September 05, 2026
🏷️ Format: Hands On Build
