Personal-assistant benchmarks move to real errands: LifeVal scores everyday chores, with best GPT-6 Astra records at 100% on several tasks
Synopsis
Evaluations of personal assistants have shifted to everyday errands. OpenAI started from the American Time Use Survey and the Multinational Time Use Study, added a demographically representative survey of 3,995 respondents to find where people most want help, and assembled the tasks into LifeVal, which it used to track model progress on such work over the past year. Among example tasks for current models, scores reach 93.1% on enrolling in the right kindergarten after a move and 90.4% on a care and meal plan for a parent newly diagnosed with diabetes; a table of best recorded GPT-6 Astra scores lists 100% for booking a sensory class, scheduling a car inspection and ordering halal groceries.
Interpretation
The yardstick for a personal assistant has moved from what a model knows to whether it can finish a single errand. Earlier assessments centered on academic tests of model knowledge, whereas LifeVal's tasks start from lists in the American Time Use Survey and the Multinational Time Use Study and are then filtered by a demographically representative survey of 3,995 respondents, covering caregiving, health, education, household administration, personal finances, shopping and government services. On individual example tasks for current models, coordinating a community's flood-preparation volunteers scores 98.0%, arranging a way to watch a game remotely with a grandparent 96.3%, and finding a home that fits a family's budget and commute 100%.
Finishing such errands takes action, not just a reply in the chat. Over the past year OpenAI trained its models on browser use, computer use and connectors tasks to be faster, more capable and better aligned, and putting those capabilities into ChatGPT widened the kinds of tasks it can help with, moving it from a chat tool toward an agentic personal assistant. The tasks are compound by construction: ordering halal groceries means placing an order for three meals inside a budget and itemizing it in a spreadsheet, and the home-finding task puts the deed in a cart for a person to review.
Perspective
The evaluation measures whether a model can complete multi-step errands in caregiving, health, education, household administration, personal finances, shopping and government services, reported as per-task percentages.
The table header reads best recorded GPT-6 Astra score, so 100% is a best per-task record rather than an average, and the example percentages belong to current models, a different accounting from the table. The year-over-year trend appears as a chart, and the summary read here supplies no figures from it, so the size of the improvement cannot be quantified.