GPT-5.6 sol post-trained luna!
Jul 9, 2026 · 5:35 PM UTC
42
90
1,178
273,824
My understanding is that you gave Sol a small task involved in the post-training process (taking a config, making small modifications to a run scheduler file, and starting a run using that config and modified run scheduler file), and it successfully completed that task in a controlled environment (and this wasn't part of the actual Luna post-training process). Could you confirm whether this is correct?
(this is very different from the conclusion I jumped to when I saw the text of your post, which was "Sol, in the real world, with minimal instruction, conducted all of the work involved in pre-training the real Luna")
5
9
151
19,174
nice! we’ve been tracking this capability for quite a while! good to see progress on this.
posttrainbench.com/
1
7
2,001
length of instructions suggests the model does not have judgement as such, and its the human providing it (and iterated on the prompt to get there).
nothing wrong w it, automation has its own value, but surely you guys being this close be able to tell the difference.
i also wonder what the gates / checks are like in a setup like this.
1
2
1,999
Surprisingly, all reasoning levels of the GPT-5.6 family achieve low maximum scores on ObviousBench.com irrespective of compute (Sol at 94%) suggesting that all 3 models are potentially underestimating the challenges of these types of questions. This was not true with 5.5.
1
3
693
