If you tag 10,000 social comments with DeepSeek V4.1 Flash, most of the model bill is output. The labels that the model writes cost 74% to 83% of the model fee. The comments that it reads cost much less. Short codes instead of full JSON, plus off-peak hours, bring the measured model fee in our Chinese and English samples from about $0.34–0.50 at peak down to $0.07–0.14.
We got these numbers from two live runs on November 9, 2026, Singapore time (SGT). From 00:16 to 00:33, 38 calls to the EveryInfra data API returned 907 comments and bullet comments in English, Chinese and Spanish. From 03:21 to 03:30, we sent 602 of the clean comments to deepseek-v4.1-flash-1017 and read the usage field of each reply. We priced the tokens with the DeepSeek pricing page. The measured model fee was 20% to 54% higher than our estimate from the doc ratios.
Two more results go against common advice. Cleaning and dedupe removed only 11.8% of rows in all three languages, so the savings are small.…