Coding models compared: Write Minecraft from scratch. Here's how they stacked up.

SubscribeSign in

I asked the top 4 models to write Minecraft from scratch. Here's how they stacked up.<br>Which frontier coding model performs best in 2026? Let's compare.<br>Jul 14, 2026

Share

Note: All four builds of the game are playable here.<br>Recently the number of frontier coding models that you can realistically use to vibe code exploded from.. 2.. to 4! Meta introduced Muse Spark 1.1 and XAI introduced Grok 4.5. Meta and XAI already had models before now, but only upon these releases are they usable in coding harnesses to achieve large projects.<br>Thanks for reading! Subscribe for free to receive new posts and support my work.

Subscribe

To compare the models I came up with this experiment. A well-defined, large scope project that does not require much clarification on the specifications: Minecraft. In this test I used:<br>Grok 4.5 via Grok Build

Meta Muse Spark 1.1 via OpenCode

Fable 5 via Claude Code

GPT 5.6 Sol via Codex

All three at their max available “thinking” settings I gave each of them the same single prompt:<br>Build Minecraft, but it’s a modern AAA title built for playing in the browser. Focus on performance, complete functionality, and at least a 5 minute vertical slice that is playable. Punch trees, make planks, build crafting table, make wooden pickaxe, mine stone.<br>I have some personal experience building voxel-worlds. There’s a lot to consider. If any harness can capture even a small chunk of this, it would be impressive. Even something as simple as generating a voxel terrain and meshing it so it visually looks like terrain is impressive, and would probably take the average programmer at least an hour or two of sustained effort. Here’s how they each performed…<br>Grok 4.5

Grok seemingly did not launch any agents and simply started one large task:<br>Scaffolding the full game systems: voxel world, player, inventory, crafting, and polished UI.

Grok finished its first turn after 8m29s .<br>After a play test it seemed like it possibly completed all the requested work. But UI bugs prevents full testing where the crafting window fell off its panel.

Grok 4.5 Minecraft World

Grok 4.5 Crafting Menu<br>GPT 5.6 Sol

Codex finished its first turn after 33m09s . It produced the fully functional vertical slice. This includes requirements for mining stone, the crafting table, stick + planks, etc. It also produced sound effects that Grok did not. However the world was very small. You can quite quickly find the edge of the world. Seemingly it produced only a single chunk. Given a second turn, I assume Sol could easily set up chunk streaming and some art improvements to reach a very playable level, difficult to distinguish from a hand-crafted game.

GPT 5.6 Sol Minecraft World

GPT 5.6 Sol Chunk Streaming Missing<br>Meta Muse Spark 1.1

Meta completed it’s first turn after 10m38s<br>Meta was the only model that produced a “block breaking” animation, and a strange looking one at that. It mostly produced the demo similar to codex. Crafting worked and there was a grok-like bug where the inventory/crafting menu weren’t visible without resizing the window. But I was able to complete the full vertical slice. However, it is the only model that allowed me to mine a stone block without the pickaxe, just more slowly. Which is in violation of the rules of the game. Obviously these are not deal-breakers, but requires further turns and following up to deliver a playable game.

Meta Muse Spark 1.1 Minecraft World

Meta Muse Spark 1.1 Block Breaking Animation<br>Claude Fable 5

Claude was the most impressive result by far, but also the most time consuming by far at a whopping 1h32m total. Claude produced everything and introduced things I didn’t specify: Soft shadows, foliage that you can mine, multiple biomes including an ocean biome.<br>It spun up 41 agents to do this work.<br>Claude interestingly spent 21m14s testing things headlessly. I never saw what the tests where but at one point it said “The review found real, serious bugs — several verified by the skeptics running actual headless simulations. Let me read the full findings list”<br>Notably, only the Fable 5 version also included a shovel tool that was separately craft-able. This version was by far the best looking as well.

Fable 5 Minecraft World

Fable 5 Crafting Menu<br>So Which is Best?

The most interesting part of this effort is the extreme difference in the TIME it took to do each prompt. When Claude produces incredible results, it’s easy to declare it the victor by comparing the output. But Claude run for nearly two hours to produce that output. Grok only ran for 5 minutes. If I followed up with 20 additional prompts to Grok between play testing, you could imagine it might come near the same level of result.<br>Initially in the experiment I had planned to run 3 prompts, all the same on all 3 models. But after seeing the results the follow-up prompts aren’t the same step along the development path. Which...