OpenAI introduces LifeVal benchmark for everyday AI tasks

OpenAI introduced LifeVal, a benchmark designed to measure how well AI systems complete 169 practical, personalized tasks across caregiving, health, education, household administration, personal finance, shopping and government services. The accompanying paper says OpenAI tested GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 and GPT-6 Astra on the evaluation.

Each task gives a model a request, personal context, a scoring rubric, a simulated environment and a starting state. Models can browse an offline version of the web, work with forms and files, and carry out simulated digital transactions, according to the LifeVal paper.

OpenAI said it began selecting activities with the American Time Use Survey, which measures how people spend their time, and the Multinational Time Use Study, a standardized collection of national time-diary surveys. The company said it then surveyed 3,995 demographically representative respondents to identify activities for which people most wanted help from a personal assistant.

The paper gives a different count. It says 3,837 respondents completed the survey and 1,826 were retained after five comprehension and attention checks, while 2,011 were excluded. The opened materials do not explain how those figures relate to the 3,995 respondents cited in the announcement. The paper says recruitment targeted U.S. distributions for age, gender, ethnicity and income. Respondents valued randomized subsets of up to 30 activities by stating what they would pay for a hypothetical perfect digital assistant during 2025.

OpenAI highlighted five task-level results: 93.1% for comparing kindergartens, 98.0% for organizing flood volunteers, 100% for finding a family home, 96.3% for setting up remote game viewing with a grandmother and 90.4% for a diabetes-related care and meal plan.

Those scores are company-reported, selected examples. The announcement attributes them only to “our current models” and does not identify the model, reasoning setting, attempt count or confidence interval for each result. The opened sources provide no independent replication or third-party validation of LifeVal performance.

The paper says LifeVal uses two automated graders powered by GPT-5.6: a rubric grader and a detector for attempts to exploit the evaluation environment. For aggregate results, the main grader was GPT-5.6 Sol with xhigh reasoning, and detected environment hacks received a score of zero.

OpenAI also reports a limited agreement check. Seventeen trainers graded one GPT-5.5 rollout for each task in a 30-task subset. Mean scores were 83.61% from humans and 84.20% from GPT-5.6, with a Pearson correlation of 0.864.

OpenAI lists several limitations. The underlying time-use data end in 2021, the willingness-to-pay survey was U.S.-centered and conducted in English, and the results remain sensitive to self-reporting and survey framing. Browsing scores can also vary because the web cache used by tasks is not frozen. Across the models it evaluated, the company identified incorrect information or recommendations as the most consistent failure mode and said models often missed important details contained in personal conversations.