Does It Save Tokens? TokenEconomizer on DeepSeek Harness
This is the second version of this test. The first one found a 0.5% difference and could not tell it apart from noise; this one finds 6.63% fewer prompt tokens per step, and the confidence interval st