OpenAI's LifeVal benchmark turns $1.7T of real-life tasks into 169 agent tests

dair_ai · x · 2026-11-05

OpenAI released LifeVal (arXiv:2611.06687), a benchmark for personalized agentic consumer tasks.

Surveyed 3,837 US consumers ranking 148 life-task categories by willingness to pay, totaling over $1.7 trillion in annual value
The most valuable tasks became 169 agent tasks: managing finances, moving apartments, job hunting, finding childcare, booking flights
Each task was written by someone who has done it repeatedly in real life, using their own ChatGPT conversations, memories and files as personal context, with a rubric, working environment, offline web copy, and three rounds of human review
Scores rise steadily from GPT-5.2 to GPT-6 Astra; the most common failure across models is a wrong answer, often from missing facts buried in past chats

More from Models