AI can pass medical exams and write complex code, but can it reliably help you plan a family budget or find the right school? OpenAI’s new research introduces LifeVal, a benchmark designed to close the gap between what AI knows and what it can do.
The Problem with AI's Report Card
For
years, the power of artificial intelligence models has been measured using academic benchmarks. These tests, often based on standardised exams for law, medicine, or mathematics, are great for showing an AI's raw knowledge and reasoning ability. However, they don't capture a crucial element: real-world usefulness. A model can score perfectly on a test but still fail at a seemingly simple, multi-step task like finding the best flight options for a holiday that fits within your budget and schedule. This is the gap that OpenAI is trying to address. Users often find that while assistants like ChatGPT are helpful for generating ideas or text, the 'last mile' of actually completing a task—filling out forms, comparing complex options, and executing a plan—still falls on them. This leaves a disconnect between the incredible potential of AI and its practical, everyday helpfulness.
Introducing LifeVal: A Test for Real Life
Announced on November 2, 2026, LifeVal is OpenAI's answer to this challenge. It’s not another academic exam but a comprehensive evaluation framework designed to measure how well AI models perform as personal assistants on practical, everyday tasks. Instead of testing abstract knowledge, LifeVal assesses an AI's ability to help with the messy, real-world chores that people actually face. The creation of this benchmark was meticulous. OpenAI started by analysing huge datasets on how people spend their time, including the American Time Use Survey and the Multinational Time Use Study. They then conducted a new survey of nearly 4,000 people to pinpoint the areas where a personal AI assistant would be most valued. The result is a test that reflects the genuine needs of users.
What Does LifeVal Actually Measure?
LifeVal focuses on multi-step workflows across seven key domains: caregiving, health, education, household administration, personal finances, shopping, and government services. The tasks are designed to be realistic and complex. For example, a task might involve finding a new kindergarten after a move, which requires comparing schools, understanding enrollment rules, and considering location. Other tasks include creating a meal plan for a parent with a new health diagnosis, organising volunteers for a community flood response, or navigating the paperwork for government services. The evaluation isn't just about whether the AI gives a correct answer; it measures whether the AI can successfully complete the entire workflow, combining web browsing, using applications, and respecting real-world constraints like budgets or deadlines.
A New Direction for the AI Industry
OpenAI revealed it has been using LifeVal internally for the past year to guide and improve its models. The results show significant progress. For instance, the GPT-6 Astra model reportedly scored 80.6% on average, a major leap from the 49.3% achieved by an earlier model in December 2025. While these scores are impressive, OpenAI is currently the only company publishing results on this internal benchmark, making direct comparisons with competitors impossible for now. However, the real significance of LifeVal isn't just the score; it’s the signal it sends to the entire industry. By creating a benchmark that prioritizes practical, agentic ability—the power to 'do' things rather than just 'know' things—OpenAI is setting a new north star for AI development. The focus is shifting from building ever-larger repositories of knowledge to creating genuinely useful assistants that can reduce the burden of daily life.
