I A/B tested GPT-5.6 Luna and GPT-6 Luna on the same engineering tasks

I ran GPT-5.6 Luna High and GPT-6 Luna High through 10 real engineering tasks on the same ~60k LOC C++/Python codebase. Each task started from the same repo state in separate worktrees with the same bounded requirements. After each task, GPT-5.6 Sol independently compared both implementations and picked one to keep. It's obviously a small sample size but here are the results: - Luna 6 won implementation speed 6-4 and used fewer tool calls. - Luna 5.6 won first-pass quality 5-3 and review quality 4-2 - Sol preferred Luna 5.6's final code 6-4, largely because its review/fix loop recovered more complex tasks. Luna 6 looks strong for narrow implementation work while Luna 5.6 looks stronger when the task has a wider integration surface. Full writeup here