Dynamic model and reasoning orchestration in Codex/Work

[OldButTalented]

This is product feedback rather than a troubleshooting request.
I believe Codex/Work should move from static model selection to dynamic runtime orchestration of model and reasoning level.
Today, the user selects a model/reasoning configuration and Codex keeps that choice during the task. This means the user has to think about model capability, reasoning level and the shared 5-hour allowance while trying to solve the actual problem.
I see that as the wrong abstraction boundary.
The user should focus on the task to be solved, not on operating the AI infrastructure.
A better execution model would be:

• start each step on the most efficient model capable of handling it
• automatically escalate model/reasoning when the current step requires more capability
• de-escalate again once the difficult part is completed
• continuously balance task complexity, quality requirements and remaining usage allowance
  This should happen during execution, not just as a one-time routing decision before the task starts.
  In effect, model choice should become an internal runtime optimization problem.
  After 30+ years working with product development, I have repeatedly seen the same principle: customers are interested in achieving their task, not in the internal complexity of the product used to achieve it.
  I think this could become an important competitive parameter for AI tools. Competition has focused heavily on stronger individual models. The next step should also be about how intelligently the platform orchestrates those models as a system.
  I do not want to spend time learning which model/reasoning combination gives the best quality-per-usage ratio. I want to spend that time on the engineering or analysis task itself.
  The best AI tool should let the user concentrate on the task. The tool should manage itself.

[Roxy373-ai]

Has anyone tried letting their ChatGPT dot choose the model and reasoning effort for each delegated task automatically, based on task complexity?

If so, what instructions or configuration do you use, and how reliable has it been compared with selecting them manually? I’m especially interested in real experiences with simple tasks versus more complex work.