I'm using whatever is free/not expensive. gpt5-mini is good enough for many tasks, but it requires patience and planning. Mid-tier models seem to hallucinate a lot. I switch to high-tier models (e.g., opus-4.6) for reviewing/fixing my own configuration of agents (soul/rituals, tools, etc).
I think, but I'm not an expert here, that the success highly depends on defining properly the main "role" of the agent and not mixing them. Have multiple agents. Project is done - dump it and start from scratch.