GPT-5.6: The Review

21:36 A review actually starts.

5.6 Sol is the first model capable of one shotting a chrome extension that replaces all mentions of obsidian with "I sniff farts"

$200k later the TLDR is it’s good 👍

For front end design the best hack I’ve found is to 1. use the image model to make the design you want. 2. Screenshot whatever shitty base design it made on its first try. 3. Give it back the image it made and your screenshot and have it describe in detail all the differences between the two and the exact changes you’d need to make to get to the good design. 4. Paste that back in a new prompt and tell it to do it. The results really are orders of magnitude better. And then you can take another screenshot and repeat that same process a few more times and it’ll get to where the image is at within a couple of prompts

I think it's the review guys

Pro tip, 5.6 is actually pretty damn good at design if you prompt it to use image generation to create design mockups to use as a reference first.

Bro gonna release Fable 5 vs GPT 5.6 after Fable access is taken from us

Honestly, i had the worst of times with 5.5, it quite literally drove me mad with its fixation on incremental fixes, disregard of the broader issue and already elaborated-on considerations, its immediate pivoting on easy user push-back and its general non-adherence. It mighta work well for executing long pre-defined tasks and implementation plans - but that's about it. Now, after picking up 5.6 yesterday, having spent 14+ hours already with it, it feels like stepping into a warm bath. Finally a friggin GPT model that shows some actual sense (again).

I used to use Matt's workflow step by step. But now since the workflow itself is already saved in memory, it just does the whole workflow in one prompt even without me asking it to. 5.6 is truly magical

I have a skill called /clean-room-loop . After Sol checks its own work, it passes empty context and just the files and check instructions to a sub agent. Whatever the sub agent finds, it fixes and runs the loop again until “clean”. I have found running: Luna low till “clean” Terra medium till “clean” Sol high till “clean” Is just faster and burns less credits. First few passes on Sol High caught obvious stuff that Luna would have caught anyway. But it was much slower. Terra catches one or two and runs clean. Sol maybe runs two passes before clean. Before that would have been 6-10 passes just using Sol.