OpenAI published LifeVal on November 2, an evaluation it says it has used for the past year to track how its models handle everyday tasks: finding childcare, budgeting, government paperwork, groceries, meal plans. Task areas came from the American Time Use Survey, the Multinational Time Use Study and a new survey of 3,995 people about where they'd want help.
The paper describes 169 tasks, run against GPT-5.2, 5.4, 5.5, 5.6 and GPT-6 Astra. Users wrote tasks from their own lives along with a rubric for each one, and the model works in simulated environments using browser, computer and connector tools, with an 8-hour timeout per run.
The grading setup is the most reusable part. A primary grader (GPT-5.6) sees the task, the rubric, the model's final answer, any files it made and a diff of the backend state from start to finish, and only gives credit for actions that actually happened. A second grader looks for hacking: if the model calls the environment's internal APIs, the run scores 0. On a 30-task check, human graders averaged 83.61% and the model grader 84.20% on the same runs.
OpenAI's page shows per-task scores like 93.1% for a kindergarten shortlist and 90.4% for a diabetes meal plan. CNN, as cited by the Artificially Intimidating brief, reports GPT-6 Astra at 80% overall against 49.3% for GPT-5.2. The paper says the most stable failure across models is wrong information, often because the model misses details sitting in a user's past conversations.