Peter Yang, who says he has no idea how to use Blender or Godot, spent one Friday night building four games simultaneously with GPT-6 Astra on the Medium model (the $20 ChatGPT Plus plan — “you really don’t need Astra high or ultra”). ~24 minutes on his own channel.

The setup: three free pieces

  • ChatGPT desktop app with the Astra model selected
  • Blender — free, open-source 3D modeling/animation tool
  • Godot — free, open-source game engine that turns those models into playable games
  • He had Astra install both apps and connect them itself: “let’s install Godot MCP and Blender MCP” — the AI then drives the apps directly through the MCP servers

The method he followed

  • Pointed Astra at an OpenAI employee’s blog post on building games with it, whose best practices were: describe the player experience first, establish a look with concept images, build repeatable play tests early, and use Blender for distinctive assets
  • Three-step loop per game: brainstorm ideas with the model → generate concept art with ChatGPT’s image model → ask AI to build and give feedback along the way
  • Each game took ~30 minutes to ~2 hours of back-and-forth; four games in parallel in one night
  • Half the prompts were lazy voice dictation from his phone while out with his family — “stream of consciousness” feedback like “the AI is pretty dumb right now”

Game 1 — Star Fox-style space shooter

  • His standard model test since GPT 5.5: compared GPT 5.5, GPT 5.6, and Fable 5 builds (all Three.js in-browser) against Astra’s Blender + Godot version
  • Astra’s version: wingmen with ChatGPT-image-generated profiles, falling-block obstacles, multiple stages, and a starship-destroyer boss whose parts break apart as you shoot it
  • Astra play-tested its own game using browser/computer use to check it was fun
  • ~30 minutes of conversation to a finished-feeling game

Game 2 — “Dust Line,” an FPS on a moving train

  • Titanfall-meets-moving-train idea; ~3-milestone plan instead of one-shot build so it could be tested along the way
  • Rockets, flamethrower, reloadable gun, jetpack enemies, boss ship at the end
  • Iterated largely via voice feedback from his phone while he was out with family

Game 3 — “Ashvall,” a Starcraft-style RTS

  • Nostalgia pick: Starcraft was his high-school obsession (he skipped classes to play it)
  • Marines, medics, and tanks he can build; siege tanks that actually enter siege mode; turrets, orbital strikes, Zerg waves, capture-the-relay base defense
  • ~a dozen prompts: “make the game too easy — add medics, three Zerg types, more aggressive enemies”
  • Quick tip he passes on: asking for better lighting effects makes 3D games look dramatically better

Game 4 — “No Moat,” a roguelike deck builder (the real game)

  • Closest to an actual game: you run an AI lab as a roguelike deck builder about an AI founder trying to build a generational company — climb a map from C-round to AGI
  • Pick a hero (the dropout, the professor, the Xoogler employee), recruit companions (designer, principal engineer, research scientist, vibe coder); enemies are scope creeper, procurement manager, reply guy, memory leak, predatory VC (“Looks like the Joker”), rival founders
  • Cards include “ship it”, “hype tweet,” “rate-limit this scope”; a GPU rack blocks attacks
  • Started in Claude (Fable) because Astra access came a day late — but Claude has no image model, so the first playable version was HTML with emojis; moved to ChatGPT once Astra was available, where the image model generated all the 2D art (he tried Blender 3D characters, but the 2D art was better)
  • ~2 hours of feedback including “check the balance against proven games in this genre”
  • He never looked at the code once; shipped it via ChatGPT Sites (remember to flip the share permission to public); playable link in the video description and first comment

Astra launch backlash — his take

  • The launch drew backlash because OpenAI did blog posts, press releases, and influencer access a week early while no normal person could get the model
  • His defense: third parties are likely involved in deciding release timing — “this may not have been entirely their fault”
  • His principle as a creator: hyping early access is unnecessary — launches should happen when the model is actually available to everybody; most models won’t change your life, so he’d rather show what he actually built

“These models have become so good that you can just talk to it like you talk to a capable game designer or developer, and they’ll just get it done.”