SOTA
Also known as: state of the art, state-of-the-art, best in class model
SOTA is thrown around constantly in model announcements, research papers, and Twitter arguments. It's not a permanent label: what's SOTA today gets beaten next month. The term only makes sense relative to a specific benchmark or task, which is why 'SOTA' without context ('SOTA on what?') is often marketing rather than information.
The most common use you'll see as a builder: model providers announcing a new release beat competitors on SOTA benchmarks like coding tasks, reasoning tests, or multimodal evaluations. Whether those benchmarks reflect real-world performance on your use case is a separate question. Many builders have found that a 'SOTA' model underperforms for their specific task compared to a smaller, cheaper model they've tuned more carefully.
Knowing the term helps you parse release notes and AI news quickly, and also helps you push back on marketing claims. 'SOTA' is a snapshot, not a verdict.