OpenAI has introduced LifeVal, an evaluation intended to measure how models perform on everyday tasks such as childcare, household administration, personal finance, shopping and access to government services. The task set draws on US and international time-use studies and a new demographically representative survey of 3,995 respondents about where people would value help from a personal assistant.
The published examples are multi-step workflows rather than isolated knowledge questions. They include comparing schools, organising flood-preparation volunteers, finding a home and preparing a care and meal plan. OpenAI reports high scores on several examples and says the evaluation has been used to track improvements in browser, computer-use and connector tasks over the past year.
LifeVal is an OpenAI-designed benchmark presented by the model provider, so the results need methodological scrutiny and independent replication. High task scores do not by themselves establish that an agent handles consent, privacy, changing circumstances or harmful mistakes safely. The useful editorial question is whether evaluation captures the whole journey: authority to act, quality of evidence, reversibility, human review and the consequences when a plausible answer produces a bad outcome.
