Morning Edition · Friday, October 16, 2026Published at 2:26 AM EDT · New York

Researchers show saturated benchmarks still carry capability signal in transcripts, and a live-clock test reverses model rankings

Pairwise judging of saturated transcripts cut capability-prediction error 5.3-fold, while RealtimeBench found that GPT-6.1-sol loses to a faster model once games keep running during inference.

Two new arXiv papers address the same problem: leaderboard scores no longer separate frontier models, or separate them in ways that do not match deployment. The first, "Evaluating Model Capabilities After Benchmark Saturation" by Morio Hama…