Module 2 · Optimising and operating prompts · Lesson 7
Model migration and regression suites
Moving a prompt to a new model safely - the API breakages, token and cost changes, and behavioural drift to check, how to design a regression suite that catches negative flips, and a worked migration of the report summariser from Claude Haiku 4.5 to Haiku 5.5.
By the end of this lesson you can
- Run a model migration in order - API changes, token recount, regression suite, prompt audit, effort sweep, shadow and cutover - and explain why each step comes where it does
- Recompute cost per call after a tokenizer and price change, including thinking tokens
- Design a regression suite with control, regression, capability and safety tiers graded on properties rather than exact text
- Compute negative-flip rate and a paired test between model versions, and decide when a migration is blocked despite a higher average
- Use the prompt changelog to find instructions written for the old model and re-test them