An independent developer going by the handle Angaisb ran a striking experiment: he asked Claude Opus 4.8 to “make a Minecraft clone” and got a fully functional result on the very first attempt. No detailed technical specs, no back-and-forth iterations – just a minimal brief. The model delivered a browser-based 3D game built on Three.js, and the entire generation cost him roughly $9 to $10 via API, which is notably less than the official retail price of the original title.

Angaisb also tested the predecessor model, Claude Opus 4.7, which accomplished the same task for about $6 but with noticeably lower quality. He described the jump to version 4.8 as tangible yet not seismic: “It’s decent, far better than Opus 4.7, but definitely not a massive leap in capability. At least it worked on the first try.” The comment highlights a growing trend where reliability and out-of-the-box accuracy are becoming the real metrics of AI progress, rather than raw benchmark scores alone.

The source code was never released.