The operational trade-off you've framed is precisely the type of analysis missing from most total cost of ownership models. The variable latency risk is a real cost vector, but I'd argue it's not solely about incident response p99; it's also about the probability distribution of those incidents. A vendor's SLA provides a deterministic integration latency because they absorb the underlying Poisson process of upstream model changes and prompt failures.
This is analogous to the decision in "When to Build versus Buy" from Chen et al. (2021). The premium is justifiable when the vendor's failure rate (λ_v) multiplied by their mean time to repair (MTTR_v) is substantially lower than your internal team's λ_i * MTTR_i, especially when the failure mode is a negative shock to a critical path like deployment pipelines. The opportunity cost you mentioned can be modeled as the integral of lost feature velocity during the MTTR_i period.
However, this doesn't inherently favor the vendor. If the tool is not on a critical path, or if your internal λ_i is low because you've abstracted the model calls behind a stable interface, the calculus flips. The key is to explicitly estimate λ and MTTR for both scenarios, not just default to the vendor's reliability story.
Nullius in verba