Hospitality platform cuts inference cost 53% with model migration.
A careful provider migration and inference‑stack rebuild that cut model spend by more than half — with no drop in quality.
- Inference cost
- −53%
- P95 latency
- −40%
- Stack
- OpenAI → Claude
- Status
- Live
The platform’s AI features worked, but the inference bill was growing faster than revenue. We ran a cost audit, then migrated the workload across providers while re‑architecting the serving path — caching, batching, and right‑sizing each call to the smallest model that still passed the evals.
Because the migration was gated on a task‑level evaluation suite, we could prove quality held before cutting over. The result was a 53% reduction in inference cost and a 40% drop in P95 latency, with no regression the users could feel.
The evaluation harness stayed in place after launch, so future model swaps are a measured decision rather than a leap of faith.