Not really that weird. This isn't intended to be a "general" model. This is a coding model so they showed the coding evals. The assumption would be relative to GPT5.1, non-coding evals would be likely regress or be similar.

Like when advertising the new airliner, most people don't care about how fast it taxis.