Many of our integration tests were written when GPT-4 was the default model.
Review and upgrade. Hopefully there's very little to do.
What should the implementer watch out for? What are the risks?