Claude Opus 4.6 vs GPT-5.4 on real-world coding: SWE-bench Verified vs Pro

SWE-bench Verified says Opus leads (80.8% vs 80.0%). SWE-bench Pro says GPT-5.4 wins by 28% (57.7% vs ~45%). Terminal-Bench gives it to GPT-5.3-Codex (77.3% vs 69.9%). Which benchmark reflects actual developer experience? Are we measuring memorization or engineering ability? Source: https://smartscope.blog/en/generative-ai/chatgpt/llm-coding-benchmark-2026/