Comment by teravor

if you are willing to put in the work you can devise a hyper specialization for which you can create an RL playground (which is likely absent in the frontier labs), then you can obtain cutting edge results on it by post training an open weights model using minimal resources.