I think we're talking past each other here. Currently all custom_jsons have to be parsed because they might be used by a hived or hivemind. What I'm suggesting is that custom_json be broken out into one kind that is parsed by core services, and another that can be used by third parties but is not parsed by hived or hivemind. The latter would presumably have lower CPU usage (almost none beyond the base to parse an op at all) and therefore could have lower RC cost.

I mean the latter doesn't even have to be json at all. It could be custom_text. We do have custom_binary (I guess this has lower cost, or it should), but people seem to prefer storing json/text rather than packing a binary.