deepseek-v4.1IMAGE
Thinking in Chinese burns more tokens: DeepSeek V4.1 Flash uses 56% more
An 8-model test of thinking in Chinese found 7 got more expensive: DeepSeek V4.1 Flash burned 56% more tokens and Claude Haiku 5.5 43% more, with no accuracy change.
AIWare
@aiwaremaxArchived 12 days agoOriginally posted 2026.10.19
AIWare @aiwaremaxOriginally posted 2026.10.19
原来让 AI 用中文思考能消耗 token ！！！ 有人拿 8 个模型实测，7 个反而更费 DeepSeek V4.1 Flash 多烧 56%，Claude Haiku 5.5 多烧 43% 准确率一点没变 那以后是不是经过一次中英转换才比较好？？？
x.com/aiwaremax/status/2111972361398579572Archived 2026.10.19 12:05