This research paper explores a new concept. The concept of a scale of harnesses. If we take openai's latest model and try to build an app with it within the chatgpt webapp, versus running a local model like qwen 3.8 27B inside of a good harness like claude code, we will be surprised to see that qwen 3.8 27B, while being probably 100X smaller, will outperform openai's model. This is because of harnesses. And Optimus Studio raises the bar, with a big focus on open weight models. We have gotten early taste of what test time compute, self evolving harness and recursive self improvement can look like. This is what's covered inside this paper ( which is my first ever paper as a solo 17 years old, hope it isn't too bad ).