GLM-5.3 Flash: 17x cheaper, trails full model by 5.6 points

Together AI ran 900 DeepSWE rollouts: GLM-5.3 Flash costs $0.24 per rollout vs $3.99 for GLM-5.3, solving 264 tasks per $100 vs 17. Flash trails by 5.6 points at pass@1 but only 2.6 at pass@4, and a cascade routing solves 80.9% of tasks at $1.70 each.