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.