Claude Opus 4.7 hits inconsistency on MineBench

Claude Opus 4.7’s performance on MineBench serves as a reminder that state-of-the-art benchmark scores do not always translate to real-world creative utility. With an average inference time of 43 minutes and costs nearing $275 per build, the model is significantly more expensive and slower than its predecessor. Its tendency to prioritize scenery over core build prompts suggests an attention shift possibly tied to its new adaptive thinking mode. While the model may be optimized for academic evaluations, it struggles with the tool-heavy, multi-step logic required for complex voxel art.