Differences Between GPT-5.6 Sol Pro and GPT-6 Astra Pro on MineBench.ai

- Every Astra build was valid on its first attempt, requiring zero retries within our harness; that reliability, alongside improved token efficiency, likely contributed substantially to lower cost.

- **Average JSON Size:** **128.53 MiB**
- GPT-5.6 Sol average JSON: 91.58 MiB

GPT-6 Astra Pro was quite a surprise. It didn't look all that great on benchmarks like Artificial Analysis, but I think its the biggest jump we've seen from a model so far. Some of the builds are at a point where they same more photorealistic or blender creations rather than created by voxels.

Average generation time stayed roughly the same as the previous generation of GPT models, but the cost looks to be much lower. GPT-5.6 Sol cost us around $700 to benchmark; the current estimate for GPT-6 Astra Pro is $34.71, though the provider dashboard hasn't updated yet (OpenAI dashboard shows $0 cost). I'm pretty sure the lack of retries is a major part of that: every build Astra outputted was valid on the first attempt, which I believe is the first time we've had a model require zero retries – usually they require at least one.

I'm not sure I agree with Greg Brockman that this model is AGI, and I haven't used it enough to weigh in on that. What does stand out is how well it seems to understand what matters in its builds. Like when to add additional scenery to a build or when it's better to stay focused on the object requested by the prompt.

Oh also this model is the most consistent with getting orientation of text correct! The only time text was backwards was the "Atlas" sign on the skyscraper build. The details in all of its builds are genuinely insane, I encourage you to try walking around the builds in MineBench ^^

- **If you enjoy these posts please feel free to help**  **fund**  **the benchmark**
- All funds are currently going directly towards API costs for benchmarking new prompts

- **Alternatively, if you have the API credits, please feel free to add prompts and generations to the gallery and post them around!**
- This is actually preferable to donations to me directly, the hosting expenses and whatnot I've always been able to cover out-of-pocket, just the API costs were hard to cover 😓

Essentially it's a benchmark that tests how well a model can create a 3D Minecraft-like structure.

So the models are given a palette of blocks (think of them like legos) and a prompt of what to build, so like the first prompt you see in the post was a fighter jet. Then the models had to build a fighter jet by returning a JSON in which they gave the coordinate of each block/lego (x, y, z). It's interesting to see which model is able to create a better 3D representation of the given prompt.

The smarter models tend to design much more detailed and intricate builds. The repository readme might help give a better understanding.