“During the development of its 'deep research' product, OpenAI designed custom evals concurrently with product development and used them to fine-tune the model, creating a continuous learning process to improve performance on hero use cases.”