Archive

OpenAI's evals lead on why old benchmarks are breaking down

As frontier models saturate existing benchmarks, the work of designing harder, more meaningful evaluations becomes the primary mechanism by which the field can track — and anticipate — the pace of AI capability growth.