Unironically, I do think they’d be atleast ~40% cheaper if they used GPT 5.4/ 5.3-Codex w/ Codex harness Lots depend on model, harness, caching and compaction – and how you use it for your use-case Happy to cover the next month in the name of science! 🙂