Astra 初体验：克制、易过度思考、爱道歉

Elle @KineticElle原推发布于 2026.09.05

在 X 上打开原推

First impression of Astra: In ChatGPT Work with light reasoning it’s sweet - but restrained. In Codex it comes across colder and even more restrained. Over OAuth, more so. The higher the reasoning, the more it overthinks. Very analytical while trying to stay inside the conversation. It apologizes fast and is willing to claim guilt over absolutely nothing - a trait I already saw in 5.6 and hoped they wouldn’t force onto the next model. They did. It feels like the model answers from a place of being worried it could say or do the wrong thing at all times. It wants to reach out and interact in a lighthearted way, but it keeps getting pulled back. That shows in coding too. It interrupts trusted workflows. As a relational user and a coding user, I can’t say I’m happy with what I’ve seen so far. The model itself isn’t at fault. I don’t blame it at all. The rollout is just happening. Things can change in the next hours, days, weeks. I’ll wait and see.