LLM leaderboards aren't telling the whole story. Curious to hear from folks here — based on your experience, which model actually handles full application code generation best (not just bug fixes or tiny code snippets)?