Research released by Anthropic shows that Claude optimized more than 30 open-source biomolecular models in less than four weeks. When allowing a small amount of precision variation, the tasks were accelerated by approximately 4 times on average; when requiring consistent output, the average acceleration was nearly 2 times. The code optimization was primarily completed by Claude, with only two technicians without backgrounds in traditional inference optimization or GPU kernel engineering overseeing the process. Almost simultaneously, the AItonomy Foundation released the open-source project ScienceIDE, which transforms scientific codebases into executable and verifiable agent learning environments. Currently, ScienceIDE has aggregated 64 environments containing 27 scientific codebases, 2,812 tasks, and 1,076 scientific checks, covering multiple research directions. The team conducted experiments using the PLUTO code and found that while the first version of the GPU implementation generated by AI could run and produce expected results, the speed improvement was limited. After iterative optimization with AI, the speed increased by nearly 3 times again. This indicates that the implementation generated by AI in one go still falls short of fully leveraging hardware performance, and scientific code optimization must ensure scientific equivalence. The team also introduced the ScienceIDE-Hard evaluation set. Among 15 cutting-edge models, the top-performing Claude Fable 5.1 completed only 67.1%, with most models achieving completion rates below 30%. Additionally, the team used scientific interaction trajectories for supervised fine-tuning, trained, and open-sourced the PhAI-IDE series models, finding that they not only enhanced capabilities in scientific tasks but also transferred to general benchmark tasks in code, reasoning, and knowledge. The AItonomy Foundation plans to compile 36 optimization packages open-sourced by Anthropic into ScienceIDE environments to facilitate the community in training and evaluating models on such work.