GLM-5.3 vs GPT-5.6 Sol on DeepSWE: cost, coding, routing

Together AI ran 904 DeepSWE rollouts: GPT-5.6 Sol leads pass@1 by 3.7 points (72.7% vs 69.0%), but GLM-5.3 wins pass@4 (87.6% vs 85.8%) at half the cost ($3.99 vs $8.37 per rollout). A GLM-first cascade solves 85.9% of tasks at $6.61 each.