OpenAI publishes LifeVal, an evaluation built partly from a 3,995-respondent survey to test AI on everyday tasks

OpenAI published LifeVal on November 2, 2026, an evaluation built partly from a 3,995-respondent survey to track AI model performance over time on practical work. The tasks include finding childcare, creating budgets and ordering groceries, according to a summary from Unite.AI.

What LifeVal covers

LifeVal covers areas where users wanted help from AI and personal agents. The survey of 3,995 respondents is one of the inputs to the evaluation's design.

How OpenAI has used it

OpenAI has used the evaluation over the past year to monitor model results and direct improvements on everyday tasks. The company's stated measurement target is practical assistance.

Details not yet reported

Benchmark scores, the models tested and the number of tasks in LifeVal are not yet reported here.

What this means

LifeVal measures models on household and personal tasks, such as childcare searches, budgeting and grocery orders, rather than on academic or coding problems. Because the tasks were drawn partly from a survey of what users wanted help with, the evaluation is tied to stated user demand. OpenAI says it has used LifeVal for a year to track results and guide improvements, so the evaluation has already been part of its internal model development.