MIT researchers have built a visual harness that lifted Claude Opus 5.0 to a perfect score on the ARC-AGI-3 reasoning benchmark. The paper, uploaded to arXiv on October 1 by Qiushi Han, Keya Hu, Linlu Qiu, Cathy Wu and Kaiming He, introduces VISTA. It lets a general-purpose model perceive environments through raw screenshots and keeps a lossless visual memory of every frame the environment returns. Flipping back through the archive does not count as a step; only real actions do. VISTA raised Claude Opus 5.0's Relative Human Action Efficiency score from 40.68 to a perfect 100. The model cleared all 25 public games and 183 levels using 57.4 per cent fewer actions than first-time human players. GPT-5.6 Sol scored 98.27 under the same harness. The team concludes the bottleneck was perception, not intelligence. Swapping the official 64×64 numeric grid for plain 512×512 screenshots alone took GPT-5.6 Sol from 13.33 to 47.32. The code is public, and the authors note their models postdate the public games.