爆料：Sol 沿用旧预训练底座，部分基准不如 5.6

Token Gremlin @TokenGremlin原推发布于 2026.09.27

在 X 上打开原推

The truth is, OpenAI keeps trying to squeeze as much as possible out of the same pretraining base, Doug, while what we really want to see are models built on the newer pretraining base, Bel. That’s probably part of why Opus 5.5 feels so “new,” while GPT-6 Sol feels more like “more of the same.” In some benchmarks, Sol is even worse than GPT-5.6 Sol, with its biggest win being price.