OpenAI on November 2, 2026 published LifeVal, an evaluation designed to measure how well its models handle the practical chores of everyday life rather than academic questions or coding. The company says the tasks cover the areas where people most want help from a personal assistant, including caregiving, health, education, household administration, personal finances, shopping and dealing with government services, and that the point is to test whether a model can actually get things done instead of only giving advice.
To choose the tasks, OpenAI started from two large studies of how people spend their time, the American Time Use Survey and the Multinational Time Use Study, then ran its own demographically representative survey of 3,995 respondents about where they would value an assistant's help. CNN, which interviewed OpenAI researcher Jenny Yang before the release, reports that a smaller group of about 100 users then wrote the tasks themselves, together with the personal context they would give the model and a rubric for judging success. Contributors had to show real expertise in the area; someone proposing a financial-planning task, for example, needed years of experience working with a financial adviser.
The tasks are meant to exercise the skills a personal agent needs: navigating websites, drawing on large amounts of context such as a user's chat history, and reaching conclusions that fit the person's needs. OpenAI's post shows worked examples with scores, including 93.1% for shortlisting kindergartens after a move, 98.0% for organizing flood-preparation volunteers, 100% for finding a home in Luzerne County and adding the deed to a county records cart, 96.3% for setting up a remote way to watch a game with a grandparent and 90.4% for a first-week care and meal plan for a parent newly diagnosed with diabetes. CNN attributes these scores to GPT-6 Astra. A separate table lists tasks where GPT-6 Astra's best recorded attempt reached 100%, such as booking a car inspection and ordering halal groceries.
The headline comparison comes from CNN's account: OpenAI's testing indicates GPT-6 Astra, released in September, scores 80% on general performance across the everyday prompts, while GPT-5.2 from December 2025 scored 49.3%. OpenAI says it has used LifeVal over the past year to track and drive improvement, and that during that time it trained its models to become faster, more capable and better aligned on browser use, computer use and connectors, the features that let ChatGPT act on a user's behalf.
LifeVal extends the logic of GDPval, the benchmark OpenAI released last year to measure performance on work tasks across industries, into household life. CNN places the release in a race among OpenAI, Meta, Anthropic and Google to build personal agents, citing Meta's Muse agent, OpenAI's Dots agent and Anthropic's Claude Duet app. Yang told CNN the aim is to find what the models are not yet good at and to give people time back by handling errands such as returning a package.
Important details are missing. OpenAI's post does not state how many tasks LifeVal contains, how the reward is calculated or who grades the attempts, whether tasks run on live websites or simulated copies, and it does not offer the task set for outside use. Only OpenAI's own models are scored, so there is no comparison with rival systems, and best-attempt figures say little about how often a task succeeds on a typical try. CNN also notes that everyday tasks are broad, depend on the information a user supplies and can have subjective outcomes.