Managed model APIs are a sensible starting point for many enterprise AI projects. They reduce infrastructure work and make strong models available quickly.
Agentic workflows change the cost profile because a single request may invoke a model many times.
An agent may plan, call a tool, inspect the result, retry, and ask another model to evaluate the answer. A workflow that looks like one user action can therefore produce dozens of background inferences.
That multiplication affects more than the token bill. It also increases latency, creates more failure points, and makes it harder to explain which model call produced the final result. The cost needs to be measured per completed task, not per prompt.
For stable, high-volume workloads, this can strengthen the case for smaller models, routing, caching, reserved capacity, or self-hosting.
Self-hosting is not automatically cheaper or more sovereign. The decision depends on utilization, data constraints, operational expertise, portability, and the cost of keeping the service reliable.
The model choice still matters, but it is no longer the only architectural question.
Can we operate the whole workflow at the required cost, latency, security, and reliability?
The durable capability lies in orchestration: measuring each task, routing it to the smallest model that can perform it reliably, and knowing when a managed service or internal deployment is the better tradeoff.