A Multilingual Cost Surprise

Your product has been priced and capacity-planned assuming an average of 700 input tokens and 300 output tokens per conversation turn, based on usage data from your (currently English-only) beta. You are about to launch in Japan and expect a meaningful share of traffic to switch to Japanese, with users writing and receiving conversationally equivalent messages to the English baseline.

1. Explain why your token-based cost and context-budget assumptions are likely wrong for the Japanese-speaking segment, even though the content of the conversations is equivalent.
2. What would you actually measure to fix the estimate, rather than guessing a multiplier?
3. Name one place besides raw API cost where this asymmetry could cause a customer-visible problem if you don't account for it.