If you read the article, the central claim is that they failed at making a more efficient model, not a larger, better, GPT-5 model.