Back to news
researchOpenAI2026-11-03

OpenAI releases LifeVal benchmark for evaluating agents on real everyday tasks

OpenAI's LifeVal benchmark draws on US and international time-use surveys plus a 3,995-respondent study to evaluate models on real everyday tasks such as childcare, health, education, finance, shopping and government services.

OpenAI published the LifeVal benchmark on November 2 to measure how AI performs on important everyday tasks, moving beyond knowledge quizzes or academic tests to assess whether models can help users decide, manage affairs and take meaningful action. The task set draws on US American Time Use Survey and international Multinational Time Use Study data, combined with a demographically representative survey of 3,995 respondents to identify the daily areas people most want help with.

Coverage spans childcare and caregiving, health management, educational preparation, household affairs, personal finance, shopping, government services, flood response, housing and moving, and remote family entertainment, with emphasis on models calling tools, accessing information and acting rather than merely offering advice. Published completion rates include 93.1% for comparing and shortlisting kindergartens after a move, 98.0% for organizing community volunteers for flood protection, 100% for finding housing within budget and commute needs, 96.3% for helping distant grandparents watch games remotely, and 90.4% for building a care and meal plan for a relative newly diagnosed with diabetes.

OpenAI said it used LifeVal over the past year to drive improvements in browser use, computer use and connector tasks, as models evolve from conversational tools into agents that complete actual tasks. The core idea is that to ensure AGI benefits all of humanity, evaluation should focus on real-life goals people care about rather than how much knowledge a model holds.

LifeVal基准评测智能体个人助理