After one week with GPT-5.6 Sol, this workflow works best for me:
GPT-5.6 Sol researches the problem and designs the architecture. GPT-5.5 writes the code. Then /goal runs the whole plan end to end in Codex.
Start it before bed. Review the result in the morning.
What's your setup? How are you combining these models?
28
7
104
11,599
Very similar except I give sol the following table and an orchestrate skill that tells it how to use these different models. Then sol manages all the workers. Also a phased approach with loop back checks. 5.5 wastes more tokens and costs than some of these lower 5.6 models
Jul 15, 2026 · 7:39 PM UTC
21