Comment by VulgarExigency

The model that is most optimized around token cost is, in fact, Chinese. DeepSeek is astoundingly cheap by default, but if you use it from Reasonix (the harness optimized around its cache), it becomes even cheaper.