Can AI handle everyday errands?
OpenAI on Nov. 2 (Nov. 3 Japan time) released "LifeVal," an evaluation that measures whether AI can handle everyday errands. It covers tasks such as finding child care, booking a vehicle inspection, handling government procedures, ordering groceries and setting up a household budget, and requires the AI to actually finish them rather than give advice. The company said it has used the evaluation to improve its models over the past year.
The evaluation spans seven areas: family care, health, education, household management, personal finance, shopping and government procedures. The tasks were selected by asking 3,995 people, in a survey weighted to match the population, where they want help from AI, based on a U.S. time-use survey and a multinational time-use survey.
Gains slow around 80%
According to OpenAI, the average score was 49.3% for GPT-5.2 in December 2025 and rose to 73.0% for GPT-5.4 in March 2026. Gains then narrowed, with GPT-5.5 at 77.4%, GPT-5.6 at 80.0% and GPT-6 Astra, released in September, at 80.6%. It is unclear whether the limit lies with the models or with the tasks.
Tasks undisclosed, grading done in-house
The LifeVal tasks have not been made public, and there are no scores for other companies' models. OpenAI created the tasks and evaluates them with its own grading system. It has also not disclosed the number of tasks, the grading criteria or whether the tests were run on real websites. CellCog pointed out that it is also an evaluation the company has spent a year using as a target for improvement.
OpenAI also publishes "GDPval," an evaluation that measures work tasks, and LifeVal extends that approach to household errands. RuntimeWire said it could serve as a training target for personal agents that handle errands on users' behalf by browsing the web, operating screens and connecting to outside services.
