AI 导读
Rohan Paul 给出数据称，GPT-4 级推理的混合 API 价格在 45 个月内从每百万 token 20 美元降到约 0.33 美元，降幅约 60 倍，快于 PC 算力和互联网带宽的历史成本曲线。但 DeepSeek 将 V4-Pro 输出价格上调 2.3× 到 4.6×，Kimi K3 则因实际 token 用量更大，每完成一项任务只比 GPT-5.6 Sol 便宜约 30%。
正文
GPT-4-class inference collapsed roughly 60× in 45 months, from $20 per million tokens to about $0.33, far faster than comparable historical PC-compute or internet-bandwidth cost curves.
But the floor is no longer moving only downward: DeepSeek raised V4-Pro output prices 2.3× to 4.6×, while Kimi K3 shows why sticker price can mislead, since heavier token use reduces its apparent discount to only about 30% cheaper per completed task than GPT-5.6 Sol.
来源：@rohanpaul_ai · x.com