[abroszka33]

Ok, I guess there are some weird benchmarks where they managed to squeeze out an exponential. But even with Google's, OpenAI's or Anthropic's own benchmark it's a linear improvement between models.