It’s quite interesting to me how (relatively) cheap it is. That’s the headline for me.
Combined with the recent math finding it’s also starting to show how general models are the way even for frontier intelligence. I would also say small/medium coding tasks is pretty much solved too (not engineering/system design etc, idea -> code in small tasks), in unison with competitive coding as a whole with the recent atcoder competition.
Claude code + fable does better with multi agent workflows than Sol + terra which means either Claude code harness is amazing or Anthropic trains the models to just be aware agentically. This is again exciting as there may come a time we can have sort of frontier harness. Claude released Claude science because clearly Claude code wasn’t built for it. Maybe, in the future , one harness does all.
Great release from OpenAI nonetheless.