Skip to content

Latest commit

 

History

History
24 lines (17 loc) · 1.14 KB

File metadata and controls

24 lines (17 loc) · 1.14 KB

Kimi K3: Command Code vs. Claude Code Harness

Same model (Kimi K3). Two harnesses.

Harness choice changes cost and time a lot, even on the same model. Claude Code didn't just use more tokens per request; it also used more turns, so total time and total tokens stacked up.

Results

Metric Command Code → Kimi Claude Code → Kimi
Turns / Requests 1 request 8 turns
Input tokens 162 ~23,000+ (heavy turns ~4,000–5,000 each)
Output tokens 1,200 N/A
Total tokens 1,362 N/A
Time 33.7 s ~78.8 s (total model time)
Cost per turn (avg) ~$0.019 ~$0.0125
Cost vs. Command Code 1x (baseline) ~5.3x
Total cost ~$0.019 ~$0.10

Combined total cost (both runs): ~$0.119

Takeaway

Command Code drove Kimi K3 to a result in a single request, with low token usage and low cost. Claude Code, using the same underlying model, took 8 turns to get there, with each heavy turn re-sending several thousand tokens of context, which pushed total session input past 23k tokens and roughly 5x'd the estimated cost versus Command Code, despite both harnesses calling the same model.