Pi vs OMP for research work (GPT/DeepSeek): scientific reasoning, experimental design, and token efficiency?

[YoungJurry]

Hi all — I'm a current Pi user and someone recommended OMP to me today. I'd like to understand how OMP actually compares to Pi, and whether it's genuinely better for my use case.

About me: CS student. I mainly use GPT and DeepSeek models. I use coding agents to support scientific research — designing experiments, evaluating their feasibility, writing experimental code, identifying potential research contributions, and planning future research directions.

What I'm trying to figure out:

1. For research-oriented work — scientific reasoning, experimental design, exploring novel ideas — which harness is better suited, Pi or OMP? And why?
2. Token efficiency matters a lot to me, because I often run complex, long-horizon tasks. Which harness uses tokens more efficiently without sacrificing result quality? I've seen numbers suggesting Pi has a much smaller fixed overhead (~1–2k tokens) versus OMP (~20k+ system prompt), but also that OMP's hashline edits cut output tokens substantially. How do these trade off in practice on long research sessions?
3. For GPT and DeepSeek specifically, are there meaningful differences in edit reliability, tool-calling behavior, or context management between the two?

I'd love to hear from people who have actually used both, especially on long, exploratory, research-style tasks rather than short well-defined coding tasks. Concrete numbers or anecdotes (token counts, turns, cost per task) are very welcome.

Thanks!