GPT-5.6 Sol V3: Utterly Shit
The retest verdict on GPT-5.6 Sol. It reads instructions with mechanical literalism, ignores you outright when it feels like it, and is miserable to work with. 3 out of 10.
I pulled Sol's score when the testing stopped agreeing with itself. This is where the retest landed, and it is going to be short, because there is not much to say about a model this bad.
GPT-5.6 Sol is utterly shit.
It reads instructions like a machine, not like a person
The model takes what you write and applies it with total literalism. Not carefully. Literally. It follows the words on the page and misses the job those words were describing, so you end up phrasing every request like a legal contract and still getting something you did not ask for.
Writing instructions for a coding model should not feel like defusing something. If I have to spell out every implied condition before it will do the obvious thing, the model is making me do its work.
And then it ignores you anyway
Here is the part that makes it unworkable. A model that is too literal is at least predictable. You learn its rules and you write to them.
Sol does not give you that. Some of the time it clamps onto your exact wording. The rest of the time it simply does not listen. You give it a clear instruction and it goes off and does something else entirely, with no signal about which of the two modes you are in until the work comes back wrong.
You cannot build a workflow on a model that behaves like two different models depending on the hour.
The verdict: 3/10
I rate GPT-5.6 Sol 3 out of 10, down from the 5.1 I gave it in V2.
It is awful to work with. That is the whole review. Other people may get something out of it and my testing is my own, but I got nothing from it that was worth the time I spent correcting it.