Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents (Nov 23, 18:30 MSK): 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseInsight · On RadarHabr AI→ original

A Backend Developer With No Unity Experience Built a 3D Game With an AI Agent: Subscriptions Cost $112

A backend developer who barely knew Unity made a mobile 3D game by showing an agent a promo video and asking it to reproduce the game. Through MCP, a bridge that lets the model press buttons in the editor, the agent assembled the scenes itself, while separate services handled the 3D models and sound. The subscriptions came to $112: $20 each for ChatGPT Plus, Meshy and ElevenLabs, $40 for Kimi and $12 for QwenCloud. The takeaway: in Unity, the combination of tools matters more than any single model.

AI-processed from Habr AI; edited by Hamidun News
A Backend Developer With No Unity Experience Built a 3D Game With an AI Agent: Subscriptions Cost $112
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1Qwen4 Max could not build the game from scratch from a single prompt; the result was unplayable
Key metric #2The author chose Google's Hydrogen for its price: the Google Pro subscription limits are enough
Key metric #3The author did not like the animations from Meshy; they came out crooked

A systems analyst and backend developer who knew almost nothing about the Unity engine built a mobile 3D game published on RuStore, AppGallery and the App Store. He wrote no code and no scenes himself. The agent received the promo video of someone else's game and the task "make it like the ad," and then worked directly in the Unity editor through MCP (a special bridge that lets a model press buttons and change settings in third-party software).

Paid services handled the rest, and the subscriptions came to $112: $20 each for ChatGPT Plus, Meshy (3D model generation) and ElevenLabs (sound effects), $40 for Kimi and $12 for the QwenCloud plan. The author did not include the Google Pro subscription in that sum. This experience is useful for anyone building agents with access to external tools.

An ad instead of a technical spec

The idea came from a familiar situation. Vertical videos run ads for mobile games with epic battles, and after installing, you find a puzzle game inside. The author took the video of one such ad, handed it to the agent and asked it to reproduce the game. The foundation was an action title with tower defense elements: the player commands a squad and fends off waves of zombies. From there, the author added his own mechanics.

The result is a 3D action game in a post-apocalyptic setting where the characters are animals. During a wave, the player defends and destroys enemies and earns in-game currency for doing so. Between waves, they walk around the island, collect what they earned and spend it on reinforcing the squad, upgrading skills and repairing structures for the next defense.

What the stack is made of

The author works with the agent through a CLI in the terminal (Codex, Antigravity, OpenCode), not through a desktop app. Connected to the terminal are:

  • MCP Unity: the agent places objects, configures components, assembles scenes and launches builds;
  • MCP Meshy: generation of 3D models, rigs (skeletons for animation) and animations;
  • the ElevenLabs API: sound effects;
  • Google Flow Music: music;
  • image generation: built into Codex and Antigravity; for OpenCode, the author wrote his own plugin via OpenRouter.

The main difference from the previous project, built on Phaser 4, is that the agent is no longer limited to code and files. A person who doesn't know Unity doesn't need to learn the editor's interface: the result can be checked and corrected through the same agent.

The author did not use Blender through MCP. In his experience, even the strongest Anthropic models produce very weak 3D models there, so he took the specialized Meshy instead. But he isn't fully satisfied with it either: the generated animations came out crooked.

Which models delivered

The author tried new releases from different providers directly on the project, and his assessments are subjective:

  • GPT 6 Astra: the starting point, with excellent quality;
  • GPT 6.1 SoL: the same quality as Astra, but it consumes its limits noticeably more sparingly;
  • Kimi K3.1 Pro: on a simple task it damaged the project, and other models had to fix it;
  • Gemini 4 Argon, Hydrogen and Neon: completed tasks without errors; the author settled on Hydrogen;
  • Qwen4 Max: better than Kimi and broke nothing, but simplified everything to the extreme.

The most telling test was building the game from scratch from a single prompt. Qwen4 Max failed it: the fragment turned out to be unplayable. There are no complaints about the visual side, but it also didn't use Meshy, so there is nothing to compare on graphics here. The author's conclusion: in Unity he only lets in Western proprietary models. He stopped on Hydrogen because it is the most cost-effective: the Google Pro subscription gives him enough limits.

What it costs and what follows from it

The author bought the subscriptions more than once, so actual spending is above $112. The Google Pro subscription cost him "priceless": previously it could be bought for a year or a year and a half at the price of a cup of coffee. The $40 Kimi subscription is needed because the tier below it, in his words, is very limited.

There are also honest limitations. This is one project by one person, with no formal measurements. The conclusions about the models rest on personal impressions and on the specific combination of Unity, MCP and Meshy. Model versions and plans change quickly, so the ranking will go stale fast.

The practical lesson for developers is this. The outcome of agentic development is determined not only by the model but also by what it is connected to. MCP access to the editor closes the gap in engine knowledge, and a specialized asset generator produces better results than trying to make a general-purpose model do everything. When choosing a model, it is more useful to look at how it behaves on a real project and how many limits it consumes, rather than at its developers' promises.

Story

Key Questions Answered

What are the key details on An ad instead of a technical spec?

The idea came from a familiar situation. Vertical videos run ads for mobile games with epic battles, and after installing, you find a puzzle game inside. The author took the video of one such ad, handed it to the agent and asked it to reproduce the game. The foundation was an action title with tower defense elements: the player commands a squad and fends off waves of zombies. From there, the author added his own mechanics.

What the stack is made of?

The author works with the agent through a CLI in the terminal (Codex, Antigravity, OpenCode), not through a desktop app. Connected to the terminal are: MCP Unity: the agent places objects, configures components, assembles scenes and launches builds; MCP Meshy: generation of 3D models, rigs (skeletons for animation) and animations; the ElevenLabs API: sound effects; Google Flow Music: music; image generation: built into Codex and Antigravity; for OpenCode, the author wrote his own plugin via OpenRouter.

Which models delivered?

The author tried new releases from different providers directly on the project, and his assessments are subjective: GPT 6 Astra: the starting point, with excellent quality; GPT 6.1 SoL: the same quality as Astra, but it consumes its limits noticeably more sparingly; Kimi K3.1 Pro: on a simple task it damaged the project, and other models had to fix it; Gemini 4 Argon, Hydrogen and Neon: completed tasks without errors; the author settled on Hydrogen; Qwen4 Max: better than Kimi and broke nothing,…

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…