OpenAI has published LifeVal, a new evaluation that grades its AI models on the practical work of everyday life rather than on exams or coding puzzles. The test covers tasks such as finding childcare, enrolling a child in school, ordering groceries and getting through government paperwork, and OpenAI says it has used the evaluation for the past year to track and improve its models. Its current models score between 90 and 100 percent on the five example workflows it published.
How the test was built
OpenAI built the task list from how people spend their time. It started with two time-use studies, the American Time Use Survey and the Multinational Time Use Study, which catalog everyday activities by demographic, time of day and season. It then added a survey of 3,995 respondents, described as demographically representative, asking where people most wanted help from an assistant.
The tasks span seven areas: caregiving, health, education, household administration, personal finances, shopping and accessing government services. That spread is the point of the test. Instead of asking a model to answer a question, LifeVal asks whether it can carry a job through to a usable outcome, using browser, computer and connector tools to act rather than merely to reply.
What the scores show
OpenAI published five example workflows, scored for its current models: finding the right kindergarten after a move, 93.1 percent; organizing flood volunteers, 98.0 percent; finding a family home, 100 percent; setting up a game viewing with a grandparent, 96.3 percent; and planning meals after a parent’s diabetes diagnosis, 90.4 percent.
Each example leaves a decision for the user, such as choosing between shortlisted schools or reviewing a plan before sharing it. The model is judged on whether it finishes the job in full, not on whether it merely answers well.
OpenAI’s chart traces the evaluation across model releases, rising from 49.3 percent for GPT-5.2 in December 2025 to 80.6 percent for GPT-6 Astra, according to secondary reports that read the chart. The gap is the headline, but it carries a clear limit: every number is OpenAI’s own scoring, on a test OpenAI built, with no independent lab rerunning the tasks.
That limit matters because an average reward score is not a share of tasks finished correctly in real life. OpenAI’s page gives no task count, no dataset, code or grader link, and no breakdown of how tasks are graded. Readers see five examples, but cannot check the test, reproduce the scores, or compare another lab’s models on the tasks.
Why It Matters
Benchmarks shape what buyers and builders trust, and LifeVal shifts the question from knowledge to usefulness. A model that tops an exam leaderboard may still fail at booking, budgeting or paperwork, while a model that completes jobs could matter more to most users. If LifeVal gains traction, labs will face pressure to show work on practical tasks rather than answers.
That pressure exists in agent testing, where the best frontier models score far below full completion on work. A SemiAnalysis agentic benchmark gave its top model just 12 percent, a reminder that hard, environment-based tests humble every lab. LifeVal differs by design: its tasks are ordinary, not expert, so a high score says more about daily usefulness than frontier capability.
Our read
OpenAI is pitching ChatGPT as an assistant that acts, not a chatbot that answers. LifeVal is the yardstick for that pitch, and its value depends on what OpenAI does next: whether it opens the task set, lets outside labs run it, and publishes failures alongside successes. Until then, the test is one lab’s progress report on its goals.
Background
OpenAI says more than one billion people use ChatGPT every week, and it wants the assistant to help with the practical challenges of everyday life. LifeVal follows its earlier workplace evaluation, GDPVal, extending the same measurement idea from office work to the chores, care and paperwork that fill most people’s days.
Source: OpenAI LifeVal page and Unite.AI report
0 Comments