Two full Codex Pro plan resets, one playable jelly slime game, and one abandoned autonomous RPG: that is what testing GPT-6 Astra cost in September 2026. The result points one way. The model is powerful, and human playtesting is what turns that power into a game worth playing.
Is GPT-6 Astra worth using for game development?
Yes, for 3D game development GPT-6 Astra is worth its higher cost, according to a hands-on video review published in September 2026 by the MattVidPro channel. The reviewer built a playable jelly slime game in Unreal Engine with it and concluded the model is the most capable language model to date for spatial, 3D work, while warning that its output quality depends almost entirely on human playtesting and specific direction.
This article is a written version of that review. Everything reported here comes from the video's first-hand demonstrations and community examples, not from independent benchmark testing. OpenAI's own documentation for Codex and the model itself remain the primary sources for official capabilities and pricing.
What is GPT-6 Astra?
GPT-6 Astra is a frontier language model from OpenAI, released publicly in early September 2026 and the first model in the GPT-6 generation, per the video. The reviewer describes it as a clear step up from the GPT-5.6 tier (Sol and Luna), especially at 3D spatial reasoning, where he says it reaches a Claude-level of spatial awareness. He also stresses the limits: it does not outperform humans in every area, and on simple tasks the difference from GPT-5.6 barely shows.
What did the community build with Astra in its first week?
Within days of release, members of the reviewer's Discord server produced a surprising range of playable and animated work, all credited to individual users on the video:
- A sci-fi racing game with a minimap, multiple laps, and later launchpads and improved flight, by user Woosh
- A PS1-inspired game with a 3D character and believable animations, also by Woosh
- A playable Black Ops 2-style zombies recreation set in a railyard, with reload animations and a points system, by user Guru
- A stylized low-poly animation of a girl on a train, plus pixel-art memes and anime music video clips, by users Godly and Guru
- A mobile-style HTML game based on the Black Ops 2 'Hijacked' map and a nighttime bedroom simulator, by Guru and Brutus
The reviewer's point in showing these is authorship: each piece looks different because each person brought a specific idea, and Astra acted as the tool that executed it.
How the jelly slime game was built in Unreal Engine
The centerpiece of the review is a 3D jelly slime game the reviewer built in Unreal Engine 5.8. You play a wobbling slime blob with physics-based movement, and the build solves several nontrivial problems: a comic-book-style see-through overlay so you never lose the character in tall grass, an arc-launching jump with a visible trajectory, and a 'physics goo' particle system that interacts with the world.
The setup matters as much as the result. The reviewer connected Codex to the official Model Context Protocol server, pulled a few free assets from the Fab marketplace, and never touched the engine's technical internals himself. He estimates the same problem-solving would take a solo developer months to learn from scratch.
He is explicit that this was not a one-prompt miracle. Every iteration was playtested, with specific issues pointed out and fixed. When the model initially refused to attempt physics particles over performance worries, simply asking it to research and brainstorm solutions fixed the implementation almost entirely. He does not claim a one-day AAA game; he calls this a small playable slice, while arguing full-scope AI-built games with a human director in the loop are now plausible.
What happened when the model built a game with no human?
To test the opposite workflow, the reviewer ran one massive goal-loop prompt for a Bethesda-style 3D RPG and did no playtesting at all. The run consumed two full Codex Pro plan resets, and he says the usage was less efficient and less worth it than the jelly game.
The autonomous RPG has genuinely impressive parts: a village with distinct buildings, a drivable steampunk vehicle with working suspension, a third-person character, footstep audio, detailed Blender-built props like a washbasin with a designed spout, and interactable NPCs with dialogue.
It also has the failures that define the experiment. Two objects occupy the same 3D space and flicker, a problem invisible to screenshot-based feedback. An NPC walks far too slowly for her animation. A door is locked from one side only because the model never tried leaving through it. A canal has no water. His diagnosis: Astra's weakness is not designing 3D spaces or using the tools; it is that the model cannot experience its own game the way a human player does. He expects this gap to shrink but calls it the biggest current difference between human and AI game building.
GPT-6 Astra vs GPT-5.6 Sol and Luna: which should you use?
The reviewer's comparison is experience-based, not benchmarked. His short version: Sol was already extremely capable and remains cheaper on the Plus plan, Luna is surprisingly capable for simple agentic work, and Astra only shows its advantage when the task is genuinely hard. On a routine job like installing game mods, he says Luna would have finished about as fast. The headline differences he found are more sub-agents used more efficiently (one game dev task spawned 16, from code review to terrain) and better 3D spatial reasoning.
| Model | Best-fit task per the review | Noted trade-off |
|---|---|---|
| GPT-6 Astra | Unreal Engine and complex 3D game development | Higher usage cost; drains Codex resets faster |
| GPT-5.6 Sol | Complex 3D spaces via Blender MCP, troubleshooting, budget work | Slightly weaker 3D than Astra |
| GPT-5.6 Luna | Simple computer tasks, installing mods and GitHub projects | Caps out on hard spatial work |
He also compares Anthropic Claude Code running Claude Fable 5.1, calling it nearly as good as Astra in areas like 3D modeling, but arguing OpenAI's Codex harness makes fewer mistakes in practice. That is his personal experience with the tools, not a controlled evaluation.
Using Astra inside Codex: speed, cost, and quirks
The reviewer recommends running Astra inside Codex, OpenAI's agentic coding tool, and says he prefers it to every competing agentic platform he has used on Mac. Codex's token caching is efficient, and on light and medium thinking modes Astra is token-efficient and can be faster than Sol; large coding projects use roughly similar token amounts to Sol.
The cost still bites. Astra is more expensive per use, and the reviewer burned through Codex Pro plan resets running it around the clock. Two behavioral quirks stand out in his account: on very wide-scope prompts the model fixates on some parts and drops others (a fixation problem, not forgetting), and it intervenes mid-task to ask about direction, which feels more agentic but annoyed him while it was controlling his computer.
Computer use improved over Sol and handles unexpected click results better, but he still finds it slow and questions why it cannot chain several actions at once without handing over full control.
Why human playtesting still decides the outcome
Across both projects, the same lesson repeats. Unsteered, Astra builds systems that functionally work and pass tests, then stops. The game worlds only became cohesive when the reviewer described the experience he wanted, such as a bug-sized goo character inside a walled garden, at which point the model produced detailed 3D pipework and a wall of individual bricks.
His conclusion is that conveying authorship, describing what an experience should feel like, and noticing what to change is becoming the most important skill for getting results from these models. The model needs native video input to catch its own visual bugs, he argues, and until then the human eye is the quality-control loop. This mirrors what other creators in the AI game-dev space, including Dev Doido do canal do youtube, demonstrate in their own walkthroughs of AI-built projects.
FAQ
- When did GPT-6 Astra come out? Per the video, GPT-6 Astra appeared publicly in early September 2026 as the first GPT-6-generation model. The video reviewing it was published on September 8, 2026, days after release.
- Do you need to know coding or Unreal Engine to build a game with Astra? The reviewer built his Unreal Engine 5.8 game without engine expertise, using Codex wired to the official MCP server and free Fab assets. He says you still need to know exactly what you want and give specific, iterative feedback.
- Is GPT-6 Astra better than GPT-5.6 Sol for everything? No. The reviewer found Sol remains the better pick for budget-sensitive work and some troubleshooting tasks, and Luna handles simple installs and mods about as fast. Astra's advantage shows on hard 3D and complex agentic work.
- How expensive is GPT-6 Astra to use? The reviewer reports a noticeable bump in usage cost inside Codex and says it drains plan resets much faster than Sol, though efficient modes and Codex's caching offset some of that. Exact pricing belongs to OpenAI's current documentation, not this review.
- Can Astra build a full game by itself? The autonomous RPG test suggests not yet. The model produced impressive 3D assets and systems but shipped visual bugs, pacing problems, and level-design mistakes it could not see without a human playtester.
Turn your own AI walkthroughs into readable articles
This article exists because a 24-minute video full of hands-on observations is hard to skim, cite, or search. If you publish the same kind of content, tool walkthroughs, build breakdowns, opinions buried in narration, you can put a YouTube URL into Skala Blog and get a transcribed, structured article draft from it. The playtesting loop the reviewer describes applies to writing too: draft first, then edit with a human eye.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits