Savings on LLMs: Chinese Model Unexpectedly Starts Writing in Chinese Characters

A developer running a bilingual news feed bot that makes about 936 LLM calls per day searched for cheaper models on OpenRouter. Out of 17 free models, only 3 were usable, and they had issues like speed or lack of JSON support. The developer then found the paid Chinese model inclusionai/ling-2.6-flash at $0.010 per million input tokens, which is 8x cheaper and 4x faster than the previously used deepseek-v4-flash ($0.098). However, after deployment, the bot started outputting Chinese characters in Russian news headlines. Testing showed 17% of short summaries and 57% of long analyses contained Chinese text. The solution was to add a CJK regex detector: if Chinese characters were found, the same request was resent to the clean deepseek-v4-flash model. After the fix, contamination dropped to 0% (0 out of 8). The developer also learned three unexpected lessons: costs are driven by news sources, not user count; free models stop working when account balance goes negative even for $0 models; and missing max_tokens parameter causes huge token reservations. The final pipeline uses ling-2.6-flash as primary, deepseek-v4-flash as fallback for CJK cleanup, and gemma:free as last resort. Monthly costs dropped from $14.8 to about $8.6, a savings of 42% rather than the expected 8x because of the dual payment for long analyses.