OpenAI員工詳解GPT-5.6 Sol五個推理層級
openai員工vaibhav srivastav解釋了gpt-5.6 sol的五個推理層級——從'light'到'ultra'——分別對應何種任務複雜度。該指南幫助用戶在準確性和token成本之間取得平衡。
openai's vaibhav srivastav released a guide for the new gpt-5.6 sol model, explaining that its 'light' and 'low' levels suit quick tasks, while 'medium', 'high', and 'xhigh' are designed for increasing complexity and verification needs. the 'max' setting extends compute time for single problems, whereas 'ultra' deploys multiple sub-agents in parallel to tackle different parts of a task simultaneously. higher levels significantly increase token usage, leading srivastav to recommend starting at lower settings and scaling up only when necessary. this system introduces complexity contrary to openai's goal of a 'no interface' experience, requiring users to carefully benchmark to find the optimal cost-to-quality ratio.