Choosing a kindergarten after a move, planning meals or organizing help for neighbors before a flood: these are among the tasks OpenAI has included in LifeVal. The evaluation looks at how AI models help with daily life, beyond answering questions and passing academic tests.
OpenAI introduced LifeVal on November 2, 2026. It focuses on tasks where people would like help from a personal AI assistant. The company says it has used the evaluation to track model progress over the past year.
What stands out to me is the focus on outcomes. Finding suitable options and respecting someone's constraints is harder than producing a generic list of tips. Still, high scores on the examples OpenAI chose do not tell us how an assistant will handle every similar situation in real life.
Where the tasks came from
OpenAI started with two studies of how people spend their time: the American Time Use Survey and the Multinational Time Use Study. Together, these provided a catalog of activities across different demographics and circumstances.
The company then surveyed 3,995 people about the activities where they would most value help from a personal assistant. OpenAI describes its survey as demographically representative. That helped it select tasks that matter to people rather than relying only on abstract examples.
LifeVal covers caregiving, health, education, household administration, personal finances, shopping and access to government services. These are areas where an answer can be useful, but where organizing information and working through several steps can matter just as much.
What the AI was asked to do
In one example, a model helps a family after a move by comparing nearby kindergartens and creating a shortlist. In the scenario OpenAI presents, the assistant organizes enrollment information and asks the person to clarify two preferences before the final decision.
Another task involves preparing for a flood. The assistant is asked to organize volunteers helping four neighbors. It produces a plan for assessing the sites and preparing for severe weather.
Other examples are more personal. One asks for a meal plan, reminders and questions for a care team after a parent is diagnosed with diabetes. Another asks the assistant to arrange a way to watch a game remotely with a grandmother, including a backup if the stream fails.
OpenAI shows scores ranging from 90.4% to 100% for the individual examples on its LifeVal page. The published description does not provide enough detail about how those figures were calculated to infer a model's overall accuracy or its probability of succeeding in real life.
What this means for users
LifeVal points to the direction OpenAI wants to take assistants: from a chat response toward completing a sequence of actions using a browser, computer controls and connected services. OpenAI says it has trained its models to work more quickly and effectively with these tools over the past year.
That does not mean every example is available to every user today. Whether an assistant can perform a particular task depends on the product, the tools connected to it and the permissions granted. The LifeVal page does not provide a complete list of availability conditions.
You can already make an everyday request more useful by including your constraints: your area, budget, deadline and any decisions you want to approve yourself. Check results that affect health, finances or safety against reliable sources, and discuss them with a qualified professional when appropriate.