The hidden invoice of a cold cache: what model migrations really cost
Swap models with a normal cache and you re-purchase your entire corpus at the new model's prices. Migration leasing is the line item that prevents the line item.
Model upgrades carry a cost nobody budgets: the cache wipe. Your corpus of answered questions, weeks of accumulated hits, assumes the old model, so the standard move is flush and rebuild. Every previously-free hit becomes a fresh model call at the new model's prices, concentrated into the weeks after launch. Teams notice the bill spike and blame the new model; the real culprit is the cold start.
Crowkis treats the upgrade as a first-class workflow instead of a wipe: canary the new model on a traffic slice, compare quality against cached baselines, then migrate entries with leasing, old answers keep serving until their replacements are verified, so hit rate never cliffs and recomputation spreads across natural traffic instead of arriving as a spike.
- 1gpt-4o cache · warm
- 2canary: slice of traffic · on the new model
- 3quality holds?
- 4migrate entries with leasing
- 5new model · cache still warm
- 6stay, nothing lost
The upgrade is a workflow, not a leap of faith.
The strategic effect is bigger than the saved spike: cold-start dread is why teams postpone upgrades, running last year's model out of cache inertia. Make migrations cheap and you upgrade the moment a better model ships, which, lately, is quarterly.
The bottom line
Count the cold start in every model decision, or better, make it stop existing. Your cache should be an asset that survives your choices, not a hostage to them.