What 7.8B Claude Opus 5.5 tokens can really build in game dev

Large language models are no longer just writing code snippets or small prototypes. With the right setup, they can now drive full game pipelines: designing worlds, generating assets, wiring gameplay, and iterating like a junior game studio that never sleeps.

This article walks through a complete experiment where Claude Opus 5.5 (often called Claude Opus Max in UI) was given a gaming laptop, a massive prompt, and access to Unity, Blender, and multiple AI APIs. After almost 8 billion tokens and roughly 36 hours of continuous work, the model produced a surprisingly polished MMO-style RPG.

Setting the challenge: an AI-built MMO RPG

The core idea was simple but ambitious: let an AI agent build a full MMO-style RPG from scratch, with minimal manual development. Instead of hand-coding systems, the human provided:

• A very detailed specification of the game world, mechanics, and aesthetics
• Existing workflows and “skills” for how to structure assets and projects
• Access to tools like Unity and Blender via MCP (Model Context Protocol) servers
• Access to external APIs for generating 2D/3D art, music, sound, and more

The AI’s job: take all of this, orchestrate the tools, and turn it into a playable game with multiple races, locations, combat, UI, and even a trailer.

Designing the mega-prompt and reference pack

This wasn’t a one-sentence prompt. The setup involved several hours of preparation to create a “mega-prompt” that described the entire game vision in detail. That included:

• Desired races and classes (6 races, each with 2 genders and some customization options)
• Four main locations, including a city and a haunted area
• The look and feel of the UI and loading screens
• References to classic MMO combat and interface patterns (similar to World of Warcraft)

Hundreds of reference images were generated and then carefully curated. Only the best ones were included in the prompt as visual guidance. The human did not build the game logic directly—only defined what the game should be and how it should roughly look.

Teaching Claude how to work: workflows as “skills”

One key to making this work was not just telling the AI what to build, but how to build it. The human provided documented workflows as “skills”, for example:

• How characters are created in parts (body, armor, accessories) and then assembled in Blender
• How to structure a Unity project, scenes, prefabs, and grayboxing
• How to iterate: build a rough version first, then refine meshes, textures, and lighting

These skills were bundled with the prompt and files in an archive, then dropped onto the laptop. From Claude’s perspective, this became a mini knowledge base for how to approach 3D character creation, environment layout, and integration between tools.

Connecting Unity, Blender, and APIs with MCP

The technical backbone of the experiment was MCP (Model Context Protocol), which lets an AI model call tools and APIs in a structured way. Claude Opus 5.5 was connected to:

• Unity MCP – to create scenes, scripts, prefabs, and playtest builds
• Blender MCP – to assemble and adjust 3D models
• External APIs like Hixville – to generate 3D models, textures, images, music, and more

With this setup, Claude could:

• Generate a 3D model via API
• Pull it into Blender for assembly or adjustments
• Export it into Unity
• Wire it into scenes, animations, and gameplay systems

This is the same kind of tool orchestration that underpins many modern AI agent setups, similar in spirit to what’s discussed in deeper dives like how newer Claude Opus versions are being used as general-purpose agents.

Letting it cook: 36 hours of autonomous iteration

Once the prompt, skills, and tool connections were in place, the model was started in an “ultra-code” style mode and left to run. The process lasted around 36 hours in total, of which ~30 hours were active work (the rest lost to a laptop crash and driver reinstall).

During that time, Claude:

• Built a grayboxed version of the world, then iteratively refined it
• Continuously launched and played the game to test changes
• Ran multiple agents in parallel (around 10 at once) to speed up progress
• Logged actions and even took screenshots when significant changes were made

The result looked a lot like a human game dev workflow: rough layout first, then repeated passes to improve terrain, props, lighting, and VFX.

Why Claude Opus 5.5 instead of other models?

The experiment compared Claude Opus 5.5 with other high-end models like GPT-6 Astra and older Claude versions. Some practical differences emerged:

• Parallelism: Opus 5.5 handled many simultaneous tasks more smoothly, which matters for large projects with multiple agents.
• Speed over long runs: Over long distances (many hours and billions of tokens), it stayed relatively fast compared to some alternatives.
• Token efficiency: It used tokens more efficiently and ended up cheaper per unit of work than some comparable models (e.g., a referenced “Fable 5.1”).

For very large, tool-heavy projects like this, those differences compound. If you’re interested in broader capability comparisons, it’s in the same family of experiments as those testing whether newer Claude Opus versions can handle complex domains like advanced maths, as in this deep-dive on PhD-level problem solving.

Token usage and real-world cost

The headline number: around 7.85 billion tokens were consumed in this run.

Most of that—about 99%—was cache reads, which are much cheaper than fresh tokens. But if you priced the whole run as standard API usage, it would land around:

• Approx. $2,175–$2,200 worth of API calls
• Equivalent to burning through one full weekly limit of a high-end Claude subscription

In this particular setup, that meant theoretically being able to run ~3 similar-scale projects in a month on a $2,000 subscription, assuming similar token efficiency.

Asset generation with Hixville API

To make the world feel real, the AI needed more than code. It needed assets: 3D models, textures, UI art, music, VFX, and more. That’s where Hixville API came in.

Using Hixville (via its MCP integration and newer API), the system generated:

• ~300 3D models using Tripor and multi-view tools
• ~600 2D images, including tileable textures for terrain, stone, wood, and other surfaces
• Additional music, animations, and some video content (smaller share of total cost)

In total, about 7,000 Hixville credits were spent, with the majority going to 3D and multi-view generation. This allowed Claude to rapidly fill the world with props, architecture, characters, and environmental details without relying on pre-made asset packs.

Recovering from crashes and project resilience

At one point, the gaming laptop crashed overnight, forcing a driver reinstall and a restart of the environment. Thanks to the way the agent and project were set up:

• Claude was able to reconstruct the Unity project from its own logs and file structure.
• No major work was lost, and the build could continue almost seamlessly.

This is an underrated benefit of AI-driven development: when the model has a full mental map of the project and a history of its own actions, it can often rebuild or repair what’s broken faster than a human could.

The final result: an AI-built MMO-style RPG

So what did the finished game actually look and feel like?

Character creation and races

The game includes six races, each with male and female variants, plus basic customization:

• Skin or garment color options
• Several hairstyles and hair colors
• Distinct backgrounds behind each character type to match their fantasy theme

Races include familiar fantasy archetypes like humans, orcs, dwarves, undead, night elves, and a bull-like “bullfolk” race. Some details, like how weapons and staves align with hands, are still rough—but the overall system works.

World and locations

The world features four main locations, including:

• A large, detailed city with streets, squares, fountains, canals, and many NPCs
• A haunted, misty area with a strong atmosphere and moody lighting
• Camps of enemies scattered across the landscape
• A starting area around a chapel or church-like structure

Weather and time-of-day effects are implemented, with the ability to change the mood and lighting across the world. Some bridges and colliders have issues, but overall layout, composition, and scale feel coherent and surprisingly polished.

UI, loading screens, and maps

The interface includes:

• A parallax loading screen with fitting music
• MMO-style action bars with multiple skills and cooldowns
• A character info window with equipment slots and stats
• A world map and mini-map that resemble classic MMORPG designs
• A quest log with quests, objectives, and acceptance/turn-in flow

Visually, the UI looks like a stylized homage to older MMORPGs, with readable layouts and consistent design language.

Combat, skills, and gameplay

Combat is tab-target or click-target based, similar to older MMOs:

• You select an enemy, then trigger skills from your action bar.
• Skills include attacks, heals, totems, and an ultimate ability.
• There’s an auto-attack system with periodic hits based on timing rules.

Visual effects for spells and abilities are surprisingly strong, with storm totems, energy bursts, and healing waves that look and feel close to classic MMO spell VFX. Animations are the weakest part—hand poses, weapon grips, and some movement cycles feel off—but the game is fully playable.

NPCs, quests, and city life

The city isn’t just a static backdrop. It includes:

• Numerous townsperson NPCs with dialogue
• Quest-givers offering missions with objectives and rewards
• Environmental logic like fountains and objects that fade when blocking the camera view
• Interior spaces like a guild hall you can enter

While the quest design is simple and some colliders are imperfect, the overall impression is of a living, explorable hub area built almost entirely by AI.

A cinematic touch: an AI-generated trailer

On top of the game itself, Claude also assembled a roughly 1 minute 40 second trailer. It showcases:

• Sweeping shots of the city, haunted zones, and landscapes
• Camera paths that highlight architecture, bridges, windmills, and the moonlit sky
• A curated sequence of scenes that feel like a genuine game trailer

This is a glimpse of where AI-driven marketing assets are heading: the same agent that builds your game can also cut together a watchable trailer from in-engine footage and generated music.

What this means for game creators

The experiment doesn’t prove that AI can replace human game designers. The game still needs:

• Better animations and polish
• Deeper quest design and narrative
• Balancing, progression, and replayability
• Human taste and judgment to turn a tech demo into a beloved game

But it does show that:

• Anyone with a clear vision and some tool wiring can get a playable, large-scale prototype built mostly by AI.
• The bottleneck is shifting from technical implementation to creativity, taste, and direction.
• Within a year or two, even the workflows and skills that had to be manually provided here may be handled automatically by more advanced agents.

In other words: the floor for what a solo creator can build is rising fast. Everyone may soon be able to “create a game,” but making a game people love will still require craft.

Key takeaways for using AI agents in game dev

If you’re considering a similar experiment, here are the main lessons:

• Invest in your prompt and references: The clearer your vision and examples, the better the outcome.
• Teach workflows, not just goals: Show the AI how you structure projects and assets, not only what you want.
• Wire in strong tools: Unity, Blender, and asset-generation APIs via MCP give the model real power.
• Expect to spend: Multi-day, multi-billion-token runs are expensive but can replace weeks or months of manual work.
• Plan for iteration: The AI will build, test, and refine in loops—lean into that instead of expecting perfection on the first pass.

This project is an early glimpse of what agentic AI plus tool orchestration can do for game development. As models and APIs improve, the gap between “idea in your head” and “playable world on your screen” will only get smaller.