Understanding the upper limit of AI's capabilities is an important prerequisite for evaluating its impact on the world. Sam Altman often talks about 'GPT has doctoral level research capabilities'. But is this a marketing tactic or a verifiable fact? Yesterday, OpenAI finally submitted its answer sheet and released a new benchmark test called "FrontierScience", attempting to use "standardized question sets+rubric scores" to make expert level scientific reasoning and research subtask abilities in physics, chemistry, and biology reproducible quantitative indicators.  Its biggest breakthrough lies not in "innovation", but in its evaluability: breaking down the originally vague scientific level reasoning into rateable rubric entries. The benchmark is divided into two parts: -Olympiad: A set of theoretical questions that benchmark the difficulty of international Olympic competitions, with short answers that can be verified and test 'hard problem-solving ability'.  -Research: A collection of 60 self-contained, multi-step research subtasks designed by PhD scientists, using a rubric quantification model to assess the level of support for research objectives (rather than the entire scientific research process).  Among various cutting-edge models, GPT-5.2 is currently the strongest, with a score of: - Olympiad：77% - Research：25%  This means that at present, AI is already a top-level 'problem-solving assistant', but it is still in the early stages of research tasks that are more open, have longer chains, and require more judgment and error correction. I have created a Chinese translation of the paper: https://(randomarea.com)/frontierscience-paper/