Only when its whole operating cost divided by accepted output is lower at the workload and quality level you actually need. A low GPU-hour rate is not a cost per task.
Record monthly tasks, input and output lengths, retries, concurrency and demand variation. Price both busy and quiet periods, including capacity needed for failures or peaks. Count review labour, engineering, monitoring, support, storage and data movement once each. Compare recurring economics separately from the one-off migration investment.
Use this definition: recurring cost per accepted task = all monthly operating costs / monthly accepted tasks. If no tasks are accepted, the metric is undefined rather than zero. Compare the same service obligation; a cheap system that abandons difficult cases is not equivalent to a system that resolves them.
An illustrative example, not a forecast: a $40,000 monthly baseline delivering 95,000 accepted tasks costs about $0.421 each. A $25,000 alternative delivering the same number costs about $0.263. A $120,000 initial investment takes eight full operating months to recover if the $15,000 monthly difference persists. Development time, transition costs and price changes can extend that period.