OpenAI introduces LifeVal to evaluate everyday tasks

OpenAI introduced LifeVal on November 2 to evaluate artificial intelligence models on everyday tasks that require personal context and actions, not just conversational responses. It covers 148 categories. The methodological study is dated November 3, 2026. OpenAI announcement · LifeVal study

How a task is evaluated

Tasks may require consulting conversations, memories and personal files, using apps, filling out forms or completing digital transactions. They run in reinforcement learning environments, controlled spaces where the model acts through a browser on an offline version of the web. Each task has a rubric, or criteria for scoring its completion. LifeVal study

The test contains 169 tasks and compares GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 and GPT-6 Astra. It uses two automated evaluators based on GPT-5.6: one examines responses, files and changes to the environment; the other detects manipulations, which receive a score of zero. LifeVal study

What its figures represent

According to the study, the willingness-to-pay survey included 3,837 people, with demographic selection geared toward the United States. Each participant rated a random subset of up to 30 categories. After applying comprehension and attention checks, the researchers retained 1,826 responses and estimated willingness to pay across the 148 categories at more than $1.7 trillion annually in the United States. These are valuations of hypothetical assistance, not observed spending. LifeVal study

The announcement publishes illustrative results for comparing preschools (93.1%), organizing volunteers before a flood (98.0%) and searching for housing (100%). These are percentages for the examples published by OpenAI, not a general LifeVal success rate. OpenAI announcement

The study identifies incorrect recommendations and omissions of personal information. Its time-use data go through 2021; the survey was conducted in English and focused on the United States. LifeVal study