I wrote about the "patent" period of frontier models a few days ago. New data point is clear: for those "little brother" models, frontier labs are increasingly losing the pricing power, thanks to fierce competition. This time around, GPT 5.6 Terra ($2/$12) is right below the Kimi K3 price ($3/$15). The pricing dynamics will again be similar to big Pharma -- the premium drugs are premium, but the last gen and generics are cheap and competitive. I won't be surprised that a natural next step for frontier labs is to open-weight those models one generation behind, because it would be easier to lock the ecosystem onto the harness + tax the inference providers than keeping up all the inference stack internally. ---- GPT 5.6 的中档和低档模型都降价了，而且和我前文所写的“专利”药品一样：越是低档的价格降低越多。这次上线才三周时间就下调价格 无论是被迫还是主动选择：前沿实验室一定会持续使用这个策略 -- 即除最强模型之外（有定价权），在中档和低档模型上，不寻求很高的 margin. 目前开放权重模型也就是比最好的模型差半档的样子，因此为保障用户锁定在生态系统，就需要中档和低档模型的定价低于对应的开放权重模型。 Terra 的 $2/$12 就正好切在 Kimi K3 的 $3/$15 下方一点。 这样外推下去的话，一年内 OpenAI /Anthropic 把上一代的中档，低档模型开放权重我都不感到惊奇：为了继续让用户为最高档的模型付费，除了模型要保持领先外，生态也很重要。为了培养生态，把已经不能够赚取高额利润的模型开放出来给社区（继续有商业协议保护）是一个理性的选择。