Claude nails PyTorch-to-JAX migration

🔥 What happened Princeton researchers dropped XStack, a benchmark that tasks AI agents with migrating entire ML repos from PyTorch/NVIDIA to JAX/TPU. Claude Opus 5 passes 75% of migrations, GPT-5.6-sol just 45%, and every other agent tops out at 40%. 💡 Why it matters ML code has been chained to its hardware ecosystem – this benchmark finally measures whether agents can break those chains. One model porting three out of four repos cleanly is a leap: a year ago agents fumbled single files, now they tackle whole codebases. ⚡ Our take 75% sounds great until you realize one in four repos still breaks. That's a non-starter for production – but the trajectory is unmistakable.