AI news ·
OpenAI builds an evaluation to measure how ChatGPT performs everyday tasks
OpenAI's LifeVal benchmark tests everyday tasks like childcare, finances, and health, with its latest models scoring above 90% on most. The eval drew on time-use data and a 3,995-person survey.
OpenAI has built an internal evaluation to measure how well its models handle everyday life tasks, and says its latest models now score above 90% on most of them. The benchmark, which the company calls LifeVal, covers caregiving, health, education, household administration, personal finances, shopping, and government services - the areas where users said they most wanted a personal assistant to take action, not just give advice.
For the more than one billion people who use ChatGPT each week, the company says a helpful assistant needs to do more than respond in chat. It should be able to complete tasks that "actually get things done," according to the announcement. OpenAI used the evaluation over the past year to track and drive improvement on those tasks.
How the task list was built
OpenAI started with two existing datasets that record how people spend their time: the American Time Use Survey and the Multinational Time Use Study. Those give a catalog of daily tasks across demographics, times of day, and seasons. The company then added a new demographically representative survey of 3,995 respondents to find out where people would most value AI help.
The result is a set of tasks that reflect the practical work of running a household and a life - from finding childcare to preparing for school to ordering groceries.
Sample tasks and model scores
OpenAI published a handful of example workflows with the scores its current models received on each:
- Finding a home in Luzerne County that fits a budget, includes a separate living unit, and keeps the commute short after school drop-off: 100%. The model reported it found one match, calculated the transfer tax, and added the deed to a cart for review.
- Organizing flood preparation volunteers for four neighbors needing floodproofing help: 98.0%. The model produced an action plan for assessing the sites.
- Watching a game with a grandparent remotely: 96.3%. The model set up a primary viewing method with a backup.
- Enrolling in kindergarten after a move: 93.1%. The model compared nearby kindergartens and organized enrollment details, flagging two preferences for the user to decide.
- Meal prep for a parent with a new diabetes diagnosis: 90.4%. The model built a first-week plan with meals, reminders, and questions for the care team.
OpenAI says training over the past year made its models faster and more capable on browser use, computer use, and connector tasks. Turning those capabilities on inside ChatGPT widens the range of work it can handle, moving the product from a chat tool toward an agentic personal assistant.
Why this matters for support, HR, and other service teams
The tasks OpenAI is measuring are the same ones that fill the queues of customer support desks, HR help centers, government service portals, and insurance claims lines. If models can plan a sequence of steps, compare options, and prepare documents for review, the near-term effect on service teams is fewer routine requests reaching a human and more work arriving pre-organized. That shifts the job toward checking decisions rather than gathering information.
Teams evaluating AI tools should ask vendors for task-level performance data, not general capability claims. OpenAI's own examples show the pattern: the model completes the legwork and hands the judgment call back to the user. Organizations that set clear review points will get more from these systems than those that treat them as a single step that replaces the process. Professionals who want to build those workflows can start with AI Agent Courses or ChatGPT Courses.