I Tested GPT-5.4 vs Claude Opus 4.6 vs Gemini 3.1 Pro on 5 Real Coding Tasks

Why I Ran This Test I use all three models daily for coding. But I've never put them head-to-head on the exact same tasks. So I designed 5 real-world coding challenges and ran each model through them. No synthetic benchmarks. No cherry-picked examples. Just everyday dev work. The 5 Tasks Refactor a 400-line Express router into a layered architecture Debug an async race condition Generate CRUD endpoints from an OpenAPI spec Document a 2000-line legacy codebase Write unit tests with edge case coverage Each task was run 3 times per model; I picked the best output. Deep Dive Refactoring - Claude Wins Claude didn't just split the code - it understood the architecture. It identified two circular dependencies I hadn't even noticed and proposed clean solutions. GPT's output was solid but missed a middleware injection edge case. Gemini got the job done but with inconsistent naming conventions. Debugging - Claude Edges Ahead All three found the race condition root cause. The difference was in th