OpenAI 发布 GPT-5.6：最小模型成本降 25 倍，速度提升 14 倍
GPT-5.6 通过优化推理流程大幅降本提速，小模型性能接近旗舰，为 AI 应用部署提供了更经济的选择。
OpenAI 今日发布 GPT-5.6 构建者指南。旗舰版 Sol 在 ARC-AGI-3 测试中得分从 13.3% 升至 38.3%，输出 Token 减少 6 倍；最小模型 Luna 在 BrowseComp 测试中以 1.33 美元成本达到 84.04% 准确率，逼近此前 GPT-5.5 的 84.36%，成本仅为后者的 1/25。同时，Ultrafast 模式预览版将 Sol 的推理速度提升至每秒 750 Token，较前代快 14 倍。
正文摘录
 新智元报道  GPT-5.6 has once again raised the ceiling of AI intelligence! Today, OpenAI released the "GPT-5.6 Builders Guide," delivering an impressive report card. On ARC-AGI-3, GPT-5.6 Sol's performance surged from 13.3% to 38.3%, while output tokens dropped by 6x. Meanwhile, the "ultra-mini" Luna also delivered a stunning performance: On the BrowseComp benchmark, it achieved 84.04%, nearly matching GPT-5.5's 84.36%, while cutting costs from $33.27 to $1.33.  Additionally, OpenAI attached a detailed tutorial on how startups can run GPT-5.6 at lower costs. ![](https…