Kimi K3 vs GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing

GPT-5.6 Sol edges Kimi K3 on single-shot quality (72.7% vs 68.5% pass@1), but Kimi K3 wins on pass@k with k>1 (89.4% vs 85.8% pass@4) and costs 64% less per rollout ($4.65 vs $8.37).

The models diverge in strengths (correlation 0.46), enabling a cascade strategy that covers 95.6% of DeepSWE tasks at lower cost than either model alone.