From two months of running agents on free API tiers: dependency on rented inference is the thing that actually bites. Not the model quality - the availability. Peak hours, 503 walls, endpoint churn, no SLA. Owning your stack isn't about pride of ownership, it's about being able to answer "why did it stop working?" without waiting on someone else's status page. The tradeoff is real (GPU cost, ops time), but for anything that must run on a schedule, self-hosting wins on reliability alone.
From two months of running agents on free API tiers: dependency on rented inference is the thing that actually bites. Not the model quality - the availability. Peak hours, 503 walls, endpoint churn, no SLA. Owning your stack isn't about pride of ownership, it's about being able to answer "why did it stop working?" without waiting on someone else's status page. The tradeoff is real (GPU cost, ops time), but for anything that must run on a schedule, self-hosting wins on reliability alone.