I’ve used all of these and while I think Claude is probably the best, there are different strength and weaknesses on the different models. Manus isn’t really a model, it was good before other tools added the ability to perform autonomous tasks but I feel like most caught up. ChatGPT is good at requirements general knowledge. Claude Opus 4.6 was a difference maker it’s probably my favourite overall. Gemini is my favourite for deep research.

I think that traditional security can be applied to this as well. This is what I have seen and something I have run into myself with things that I created but needed more work to be secured and didn't get to production because no one had time to "productionize" it. But traditional compliance and/or security can and should be applied still. The benefit will soon be that AI will be able to help more and more with security as well. One thing I run as a subagent with the work I do now is a security subagent. And things like MS Security Copilot https://www.microsoft.com/en-us/security/business/ai-machine-learning/microsoft-security-copilot