Min Choi rounds up Grok 4.5 builds, including a playable FPS and UE5 cyberpunk street

• Grok 4.5's first creator wave is code-to-playable-fast, with an FPS under an hour in minchoi's FPS post and a UE5 cyberpunk street at 36 minutes for about $12 in minchoi's UE5 example.
• Web designers got a reusable workflow: viktoroddy recorded a 12-minute Cursor + Grok 4.5 website tutorial in viktoroddy's tutorial post and shared the prompt pack through viktoroddy's prompt post.
• Access surfaced through Grok Build and Cursor, with limited free CLI access claimed by ozansihay and $2/M input, $6/M output pricing claimed by minchoi's pricing post.
• Agent builders are testing Grok 4.5 for tool-using workflows, and gregisenberg claimed Hermes/OpenClaw runs landed around $2.49 per task versus about $12 for Fable in Claude Code.
• The visual-quality read is mixed: om_patel5's 3D bake-off put Claude Fable 5 ahead on a browser 3D scene, while Grok 4.5 used fewer tokens.

The demos are physical: shooters, home planners, UE5 streets, floating islands, Framer games. minchoi's space-battle example says a two-line prompt produced a playable game with sound, while tranmautritam used Grok for a Liquid Glass Button inside Framer AI Agents. The useful bit for creators is the spread across surfaces: Cursor, Grok Build, Framer, Hermes, OpenClaw.

viktoroddy's angle was workflow, not leaderboard theater: a 12-minute Cursor tutorial for building premium websites with Grok 4.5, followed by a post giving viewers the prompts.

The tutorial link itself sits behind viktoroddy's video link. The same account had just posted a longer GPT 5.6 Sol website tutorial in viktoroddy's GPT 5.6 post, which makes Grok 4.5 look less like a novelty demo and more like another model in the premium-site prompt stack.

minchoi's roundup centered on playable outputs, not static screenshots.

• A playable FPS that minchoi said was built in under an hour.
• A playable space battle that minchoi said came from a two-line prompt and included sound.
• A 3D game with rigged enemies, a map, and AI logic that minchoi attributed to two prompts.
• A polished 2D side-scroller that minchoi said landed in under 30 minutes.
• An Elon Musk management sim that chrisfirst built with Grok 4.5, including product research, server upgrades, and company upkeep.

Fast game prototyping is the cleanest creative story here. The claims are all small enough to test, and varied enough to matter.

The strongest non-game examples moved into environments, floor plans, and browser 3D.

• UE5 cyberpunk street: minchoi put the build time at 36 minutes and cost at about $12.
• Home planning: minchoi said Grok 4.5 designed a full 2D + 3D home planner in under a minute.
• Repeat home-planner claim: minchoi's later post repeated the same 2D + 3D planner example later in the thread.

The floor-plan demo is the most obvious bridge to creative tools: layout, regenerate, preview, then hand-edit.

Grok also showed up in Framer Agent experiments, where the output is closer to a deployable interaction than a coding benchmark.

tranmautritam used Grok to make a Steady Hand game with Framer Agents in his game post. A second prompt in his Liquid Glass Button post asked Grok to create the kind of glossy UI component currently dominating AI-design feeds.

The clearest access claims came from Grok Build, Cursor, and terminal screenshots.

• Grok Build CLI: ozansihay said xAI integrated Grok 4.5 into Grok Build and allowed limited free use with any X or Grok account.
• Install path: ozansihay's post included a one-line curl installer, with the linked install URL attached.
• Terminal mode: icreatelife's screenshot showed a Grok 4.5 high session with an always-approve status.
• Pricing and surfaces: minchoi's pricing post said Grok 4.5 was available in Grok Build, Cursor, and the SpaceXAI console at $2/M input tokens and $6/M output tokens.
• Regional limit: minchoi's access note said it was not available in the EU yet.

gregisenberg framed Grok 4.5 as an agent model for Hermes and OpenClaw, not just a chat model for one-off prompts.

The cost claim was the hook: gregisenberg said Grok 4.5 was more than 60% cheaper than Opus 4.8 and landed around $2.49 per task versus about $12 for Fable in Claude Code. The Hermes screenshot in ozansihay's post showed Grok 4.5 Med listed alongside GPT-5.6 Terra and Luna options.

om_patel5 posted the most useful counterweight: four models got the same prompt, generate floating-island versions of Paris, London, and New York entirely in the browser from scratch.

The ranking in om_patel5's comparison put Claude Fable 5 first, GPT 5.6 Sol as solid but rougher, GLM 5.2 as over-bloomed, and Grok 4.5 as the weakest visual result. The same post said Grok 4.5 used a fraction of the tokens the others burned, which matches the efficiency angle in minchoi's Artificial Analysis chart post.