Move prompts between models
Links on this page to features marked Preview describe designs that are not available yet.
A prompt tuned for one model often works on another, but not always as well. Switching models is a change to test, not a configuration flip.
What tends to change
| Area | What to watch |
|---|---|
| Length | Newer or larger models are often more concise or more thorough by default. |
| Instruction following | A model that follows instructions more literally may now obey a rule you added years ago as a workaround. |
| Format | JSON key order, Markdown use, list styles. Use structured outputs where format matters. |
| Tool use | How eagerly tools are called, and whether several are called in parallel. |
| Refusals | Borderline requests may be handled differently. |
A safe process
- Run your evaluation set on the old and new model with the unchanged prompt.
- Read the differences, not just the scores.
- Remove workarounds that only existed for the old model, one at a time, re-running each time.
- Only then add new instructions for genuine gaps.
- Shift traffic gradually, watching cost per request as well as quality.
Keep prompts model-neutral where you can
Describe the task and the desired result rather than tricks aimed at one model. Neutral prompts survive model upgrades with fewer surprises.
See Prompting each model, Model cards and Deprecation policy.
Was this page helpful?