“Data collected for reinforcement fine-tuning will remain useful for future, more capable models, making it a durable investment that complements model scaling.”