OpenAI unveils LifeVal, shifting AGI evaluation from exams to real-life tasks
OpenAI has introduced LifeVal, a new benchmark designed to measure whether advanced AI systems can handle the kinds of personal, high-value tasks people actually care about, rather than just scoring well on exams or academic contests. According to the report cited in the source article, the benchmark was built after OpenAI reviewed time-use studies across 25 countries and surveyed more than 3,800 consumers in a sample aligned with U.S. Census demographics. From that process, the company selected 148 consumer tasks and estimated that U.S. consumers would be willing to pay a combined $1.7 trillion per year for a perfect assistant that could complete them. The article says LifeVal focuses on tasks such as finding childcare, accessing government services, managing personal finances, hiring a nanny, booking flights with credit card points, and organizing medical records. OpenAI also released a 16-page paper and disclosed LifeVal results for GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6, and GPT-6 Astra. In the reported results, GPT-6 Astra was described as OpenAI’s most aligned model so far and showed stronger stability on tasks involving private user context and multi-step web actions. The article also highlights a broader tension: AI assistants may reduce administrative burden and save time, but they also require access to deeply personal data and decision-making workflows.


