How to Build 3D Games with GPT-6 Astra and Codex
A grounded look at building playable 3D games in Unreal Engine using GPT-6 Astra and Codex, from workflow basics to real limits.
What is GPT-6 Astra and why does it matter for game development?
GPT-6 Astra is the latest model variant that creators have started using inside Codex to generate 3D game environments, characters, and mechanics from natural language prompts. What sets it apart from prior models (referred to as “Soul” in early creator feedback) is a sizable jump in 3D spatial reasoning. People building with it report it placing objects, designing meshes, and handling physics in ways that previously required real Unreal Engine or Unity experience. That matters because it lowers the floor for who can prototype a playable 3D game. You no longer need to know how to code or navigate a game engine’s interface to get a working build. You need to know how to describe what you want, clearly and specifically, and how to react to what comes back.
TL;DR
- GPT-6 Astra shows a real jump in 3D spatial awareness compared to earlier models, handling object placement, mesh design, and physics with noticeably more competence.
- Codex connects to Unreal Engine through an MCP server, letting the model generate assets, write logic, and iterate on a live project rather than just producing static code snippets.
- Human playtesting in the loop produces dramatically better results than letting the model run unsupervised for hours, based on side-by-side comparisons of a supervised slime game versus a mostly autonomous RPG demo.
- Specific, detailed prompting outperforms vague direction. Telling the model exactly what a wall, a character animation, or a UI element should feel like gets you closer to an “authored” result than open-ended requests.
- The model can solve genuinely hard technical problems, like implementing physics-based goo particles or a teleporting camera view, when pushed to research and iterate rather than accepting its first attempt.
- Current limits include 3D flickering and z-fighting issues that are hard to catch from static screenshots, since the model still lacks native video input to self-review its own gameplay.
- Dialogue and pacing can feel mechanically generated unless you explicitly direct tone and context, since the model doesn’t inherently track how “long” narrative beats should take to land.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How does the Astra and Codex workflow actually work?
The basic setup pairs Codex (the coding agent) with the GPT-6 Astra model, connected to Unreal Engine 5 through an MCP (Model Context Protocol) server. This connection lets the AI agent do more than write isolated scripts. It can generate 3D assets, place them in a scene, wire up gameplay logic, and respond to feedback about a running build. Some workflows also bring in free asset packs (referred to as “fab assets” in Unreal’s ecosystem) to give the model existing building blocks rather than forcing it to generate everything from scratch.
The practical loop looks like this: you give the model a specific creative goal, it builds or modifies the scene, you play the result, and you report back exactly what’s wrong or what you want changed. This is repeated, sometimes across many sessions, until the demo matches your intent. One creator described building a 3D “jelly slime” game this way, where the player controls a physics-based blob character with squash-and-stretch movement, arc-based jumping, and a teleporting camera view that fades through obstacles to keep the character visible.
Why does human playtesting matter so much?
The clearest lesson from early Astra experiments is that supervision changes outcomes. In one comparison, a slime-based platformer was built through constant human playtesting: the creator played every iteration, pointed out specific problems (a camera view failing to track the character, goo particles causing performance drops, a garden environment that felt generic instead of intentional), and asked for targeted fixes. That project came out cohesive and, by the creator’s account, close to shippable quality for a small vertical slice.
A second project, a Bethesda-style RPG with villages, NPCs, and a drivable steampunk vehicle, was built almost entirely without playtesting: one long prompt, left to run across multiple sessions. The result still had real production values (footstep sounds, a working vehicle with suspension, an NPC with dialogue trees, a fully modeled tool chest and stove), but it also had noticeable problems: two 3D objects occupying the same space and flickering as a result, an NPC that walked at an oddly slow pace, and dialogue that jumped into plot details without setting up context first. None of these are catastrophic, but they’re the kind of thing a five-minute playtest catches immediately and an unsupervised run does not.
The takeaway isn’t that autonomous generation fails. It’s that the model currently can’t watch itself play. It can’t yet see a flickering texture or a character moving too slowly on video and self-correct. Until that kind of video feedback loop exists, a human eye in the loop remains the fastest way to catch and fix issues that text-based review misses.
What can Astra actually build right now?
Based on demos shared by early users, the range is wide. Creators have built:
One coffee. One working app.
You bring the idea. Remy manages the project.
A racing game with a sci-fi interior track, complete with a mini-map and multi-lap structure, plus custom spaceship designs. A zombie survival shooter modeled after a well-known arcade mode, including reload animations and a point-based weapon system. Stylized 3D animations and short films, ranging from low-poly character pieces to anime music videos built from user-supplied clips. A physics-driven slime platformer with squash-and-stretch character movement and interactive goo particles. An RPG-style village environment with a drivable vehicle, NPC dialogue, and hand-modeled props like stoves and wash basins.
The common thread across all of these is that the ideas come from the person prompting, not the model. The creative direction, the reference points, the specific details that make something feel “authored” rather than generic, all of that still comes from a human with a clear vision. Astra’s role is closer to an extremely capable, fast collaborator than an independent creative director.
Is building games this way worth the effort?
For prototyping and solo experimentation, yes, with caveats. The speed at which a single person can go from an idea to a playable, textured, animated 3D scene is not something that was practical before models with this level of 3D spatial reasoning. A solo developer without engine experience getting this far in a matter of hours or days is a meaningful shift from the months such polish would normally take.
But it’s not free of friction. Unsupervised generation can burn through significant usage (one creator noted needing two full plan resets on Codex’s usage limits for the RPG demo) without producing results as clean as a supervised, iterative session. And it’s not building finished, ready-to-ship games. What’s coming out are vertical slices and demos, not full-scope, tested, balanced products. The gap between “impressive playable demo” and “shippable game” still involves the same production concerns AI hasn’t solved: balancing, QA across edge cases, performance optimization at scale, and narrative pacing that respects player context.
The realistic framing is that Astra plus Codex is a genuinely capable prototyping and production tool for people who know what they want and are willing to direct it closely, not a way to skip the game design process entirely.
Frequently Asked Questions
Do I need to know how to code to build games with Astra and Codex?
No. The workflow described by early users doesn’t require traditional coding or Unreal Engine expertise. What it does require is the ability to describe, clearly and specifically, what you want built and what’s wrong when you test it.
What engine does this workflow use?
The demos referenced here were built in Unreal Engine 5, with Codex connected through an MCP server that lets the AI agent generate assets and logic directly inside the engine.
Why does unsupervised generation produce worse results than supervised sessions?
Because the model currently can’t see or play its own output the way a human can. It lacks native video input, so subtle problems like object flickering, mistimed animations, or pacing issues in dialogue often go unnoticed unless a human explicitly points them out.
Can Astra generate entire full-scope games on its own?
Not currently. It can produce complex, detailed, playable demos and vertical slices, especially with steady human direction, but full-scope, polished, tested games remain out of reach for a single prompt-and-wait approach.
What’s the biggest factor in getting good results from Astra?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Specificity. Being precise about what you want a scene, character, or mechanic to feel like, and giving detailed feedback after each iteration, produces noticeably better and more “authored” results than vague or open-ended prompting.
