VISTA turns Claude Opus 5.0 into a perfect gamer

🔥 What happened MIT and Google DeepMind researchers dropped VISTA, a visual harness that gives multimodal models lossless visual memory and active retrieval. On ARC-AGI-3, it pushed Claude Opus 5.0 from a 40.68 to a perfect 100.00 Relative Human Action Efficiency score. 💡 Why it matters The model cleared all 25 public games using 57.4% fewer actions than first-time human players. No retraining, just a smarter interface — the reasoning was already in the weights, it just needed the right glasses. ⚡ Our take If you're still training models in 2026 instead of building harnesses, you're burning cash. VISTA proves the real product is the environment, not the network.