Real numbers on switching-cost economics

Frontier model APIs can look expensive when you focus on the price per token. The harder question is what it costs to leave them. A migration writeup from Opus to self-hosted Ollama puts real operational detail around that question, rather than treating self-hosting as a simple swap of one model endpoint for another. You can read the original report for the full account.

For an owner, this is a useful comparator when someone proposes moving workloads from a managed provider to an in-house model. The managed API bill is only one line of the decision. You also need to account for the work of adapting prompts, testing output quality, running infrastructure, monitoring reliability, and dealing with the cases where a smaller or local model does not produce the same result. Those costs can be worth paying, particularly where usage is high or data requirements are strict. But they need to be measured against actual savings, not assumed.

The practical lesson is to avoid making this a broad “cloud versus self-hosted” debate. Pick a specific workflow, establish its current cost and quality level, then test an alternative model against it. Track the total cost to operate, not just inference. This is the kind of thing we build into an AI command centre, giving teams a clearer view of model cost, performance, and where a change is genuinely justified.