We audited Qwen4-Max and GPT-6.1 Sol on five synthetic tasks. Compare exact API costs, failed outputs, self-review results ...