In a Warp internal benchmark, GPT-5.6 Sol achieved 95% correctness at an average cost of $6 per task, while GLM-5.3 Flash achieved 93% correctness at $2