Knowledge base re-embedding
Fine-tuning re-run
Engineering & testing time
Current vs new monthly spend
| Cost component | Amount | % of total |
|---|
Vendor lock-in is mostly invisible
The API price gap between providers is easy to see. The migration cost is not. Most teams underestimate by 3–5× because they only budget engineering time and forget:
- Embeddings are not portable. Every document needs to be re-embedded in the new model's vector space.
- Fine-tuned models don't transfer. Training data, hyperparameters and checkpoint — all re-run from scratch.
- Prompts need rework. System prompts tuned for one model often perform poorly on another.
- Regressions take time to find. QA on LLM outputs is slower than QA on deterministic code.
This doesn't mean staying put forever — just that the payback period is real and should be calculated before committing. Compare providers on LLM price comparison. If you only want to hedge, see AI model router cost.