> has actively hallucinated details

Gemini models always suffered from this. They were almost always best in terms of intelligence, but one of the worst to use in practice, because they have a lot of failure modes on simple tasks that simply give factually incorrect answers, or completely break and output gibberish.