I switched to GLM-5.3 flash on high for this reason. Too many "but wait" in the Deepseek-v4 reasoning.