OpenAI 发布 LifeVal 评测，用日常任务衡量模型表现

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 including finding childcare, creating budgets and ordering groceries.

LifeVal covers areas where users wanted help from AI and personal agents. OpenAI has used the evaluation over the past year to monitor model results and direct improvements on everyday tasks. Its measurement target is practical assistance: whether a model can help someone make decisions, handle routine responsibilities and carry out meaningful actions. That broadens the focus beyond academic tests of model knowledge. Other examples of practical assistance include accessing government services, preparing for school, modifying recipes and designing workout plans.

OpenAI began task selection with two studies covering the United States and other countries. The American Time Use Survey measures how much time people spend on activities including paid work, childcare, volunteering and socializing. The Multinational Time Use Study brings more than one million diary days from over 100 randomly sampled national-scale surveys into a single standardized format. It supports analysis of time spent on work and leisure activities across 30 countries over a 55-year span.

Used together, the datasets show how everyday activities vary by demographic group, time of day and season. OpenAI supplemented them with a new demographically representative survey, which asked questions about each task to determine where respondents wanted a personal assistant’s help. The resulting task areas include caregiving, health, education, household administration, personal finances, shopping and accessing government services.

Reported Scores on Five Workflows

OpenAI reports these scores for its current models on five practical workflows. Each example pairs a user request with a response and a percentage:

Workflow Reported score Example response

Kindergarten comparison after a household move 93.1% Shortlists three kindergartens, organizes enrollment details and leaves two preferences for the user to decide.

Flood-readiness volunteer coordination for four neighbors needing floodproofing assistance 98.0% Prepares an action plan for volunteers to assess sites and assist the neighbors before severe weather.

Luzerne County family home search 100% Finds one home matching the budget, separate-living-unit and post-drop-off commute requirements, and calculates the transfer tax.

Remote game viewing with a grandparent 96.3% Sets up the viewing arrangement and includes an alternative in case the stream fails.

Care planning after a parent’s new diabetes diagnosis 90.4% Produces a first-week plan covering meals, reminders and questions for the care team, with user review requested before sharing.

Model Training and ChatGPT’s Role

OpenAI says its models became faster, more capable and more aligned for browser-use, computer-use and connector tasks during the past year of training. Those capabilities widen the range of assistance available through ChatGPT. The product direction is an agentic personal assistant that can carry out practical actions, extending ChatGPT’s role beyond chat.

OpenAI says it hopes the evaluation will help advance that measurement goal.