tl;dr: Ploy migrated their website-building AI agent from Claude Opus 4.8 to GPT-5.6, achieving 2.2x faster builds and 27% lower costs, but the switch required significant engineering work beyond swapping SDK calls. Key gotchas included: GPT-5.6 hallucinating values for all optional tool parameters (fixed with nullable schemas), OpenAI's redesigned caching requiring explicit per-workspace cache keys to avoid 0% hit rates, and reasoning replay needing `store: false` to prevent server-state errors. Their eval harness itself was also silently biased toward the incumbent model's behavior.