Cheap tokens, expensive interruptions
Compare the cost of a finished task, including the repairs.
Retrospective reconstruction of the week beginning 2025-01-06, written on 16 November 2026. This is not a claim of publication at the time.
DeepSeek V3 was barely a week old at the start of January. Model cost was an obvious thing to compare. The harder number is what it costs to get an answer you can actually use.
Imagine a batch of twenty descriptions. One setup is cheaper per request but leaves six descriptions needing repair. Another costs more and leaves one. The first invoice looks better until someone has to sit down and fix the batch.
Count that work. Count retries, checks and interruptions. Add the storage and ordinary software needed to keep the thing running. A solo builder’s afternoon is not free just because it has no line on a cloud bill.
For a fair comparison, use the same inputs and decide what counts as acceptable before running them. Keep the prompt fixed. If you improve it halfway through, rerun both options.
The result may still favour the cheaper model. Great. Now there is a reason to choose it beyond a price screenshot. Keep the comparison small enough to repeat when the model or workload changes.