Ask HN: How are you productive with GPT 5.6 Sol?

[hk__2]

I recently switched from Opus/Fable to GPT 5.6 Sol, and so far I’ve found it so dumb that it often makes me lose time. I can ask Opus/Fable one thing and I know it’ll take some time but will do it almost perfectly. With GPT I constantly have to steer it; its investigations almost always end up with overreached conclusions after minutes of looking in random places unrelated to the issue; it adds a lot of useless defensive code just like previous 5.x models. I do use AGENTS.md; I have a couple of MCPs; I enabled memories.

I see people happy about GPT 5.6 Sol everywhere, so there must be something I’m missing. If you’ve tested both model families, what have been your experience so far? Have you noticed that some things that work well with Claude don’t with GPT or vice-versa?

[runjake]

I have the opposite perspective. I primarily use Codex, but jump into Claude and Claude just feels dumber, generates worse code, and less productive.

I don't feel like Claude is awful or anything, just that Codex and it's models are significantly better.

[ga6840]

the first a few days are very good. after one week, it becomes dumber in my case. I've experienced similar thing from openai models, my guess is openai's "reasoning efforts" even in the case of Sol, is about different levels of quantization.

[hk__2]

It turns out that my issue was using high/xhigh. Going back to medium or even low made it 'intelligent' again.