I tried it, and it's good but I feel like the synthetic data used for training 3.1 does not hold up to gpt4o prob using human-curated data.