A benchmark built from time-use surveys
OpenAI published LifeVal on 2 November, an internal evaluation it says measures how well its models handle “the tasks that matter in everyday life” — finding childcare, reaching government services, building a budget, modifying recipes, ordering groceries.
The task list was not invented. OpenAI started from two long-running statistical surveys of how people spend their time, the American Time Use Survey and the Multinational Time Use Study, then ran its own demographically representative survey of 3,995 respondents asking where people would most want help from a personal assistant. The resulting tasks cover caregiving, health, education, household administration, personal finances, shopping and access to government services.
The scores it chose to show
The published page gives five example tasks with a percentage attached to each. Enrolling a child in the right kindergarten after moving: 93.1%. Organising volunteers to floodproof four neighbours’ homes ahead of a flood: 98.0%. Finding a home in Luzerne County within budget, with a separate living unit and a short commute, and adding the deed to a county records cart: 100%. Setting up a remote way to watch a game with a grandparent: 96.3%. Building a first-week care and meal plan for a parent newly diagnosed with diabetes: 90.4%.
OpenAI says its “latest models score very highly on many LifeVal tasks”. It does not say which model produced these five numbers, how a task is scored, what counts as success, who judged it, or how the five examples were picked out of the full set.
What is not published
LifeVal is not an open benchmark. There is no dataset release, no paper, no leaderboard and no independent run. Every figure on the page is a vendor-reported result about a vendor-built evaluation, and the page is explicit that OpenAI has been using it internally “to track and drive improvement in model performance on everyday tasks over the past year” — in other words, as a training target as well as a measurement.
That is not disqualifying; labs build internal evals constantly. It does mean the numbers say something about OpenAI’s own progress against its own target, and nothing comparable about any other system.
Why OpenAI is publishing it now
The framing connects the eval to capability work rather than to chat quality. OpenAI says that over the past year it has trained its models to be “much faster, more capable, and more aligned” on browser use, computer use and connector tasks, and that enabling those in ChatGPT expands what the assistant can be asked to do. LifeVal is the scorecard for that shift — from answering questions to taking actions on someone’s behalf.
What to watch
The useful next step would be the one OpenAI has not taken: releasing the task set, the grading rubric and the full score distribution, so that a result on LifeVal means something to a reader who does not work at OpenAI. Until then, a 100% on finding a home is a claim, not a measurement anyone else can check.
