Switching Costs: Migrating Between Frontier Vendors

Two frontier models can produce similar outputs from quite different prompts. Migrating from one to another usually means rewriting prompts, re-evaluating tool schemas, and re-tuning safety language. Budget weeks, not hours.

Where switching costs hide

1. 1Prompt phrasing — what works on Claude reads weird to GPT and vice versa
2. 2Tool / function-calling schemas — slight syntax differences cause silent breaks
3. 3Streaming format — server-sent events look different across vendors
4. 4Refusal patterns — the new vendor refuses things the old one accepted
5. 5Token tokenization — the same prompt costs different amounts on different models

Applied exercise

1. 1Estimate your switching cost: weeks of engineer time to fully migrate
2. 2Multiply by your loaded engineer rate to get a dollar figure
3. 3Compare to the savings or capability gain from switching
4. 4Decide if a partial migration — only the highest-volume endpoints — is the right move

The big idea: vendor switches are projects, not button presses. Plan accordingly and use abstractions where they pay off.