GDPval

GDPval measures how well a model handles the kind of tasks people are actually paid to do: drafting documents, working through spreadsheets, pulling together research, multi-step jobs with messy inputs. It was built by OpenAI as a counterweight to benchmarks made of exam questions, which reward a narrow kind of cleverness.

Results are reported as Elo points rather than a percentage, so the useful reading is the gap between models rather than the absolute number. Since OpenAI both created the benchmark and competes on it, it is worth treating the scores as one input among several rather than a neutral verdict.