News · SI
OpenAI's LifeVal tests whether AI agents can actually run everyday errands
OpenAI has published LifeVal, an internal evaluation of how well its models finish everyday tasks like choosing a kindergarten, planning meals or getting through government paperwork. Tasks come from time-use studies and a 3,995-person survey, but the task set and grading aren't public.
OpenAI has published LifeVal, an evaluation of how well AI models handle the practical jobs of everyday life, from finding childcare and accessing government services to modifying recipes and ordering groceries[1]. The company's framing is blunt: to be a truly helpful personal assistant, ChatGPT has to take actions that get things done, not just offer advice[1].
01 / 04
Where the tasks come from
Rather than invent a task list, OpenAI started with two large studies of how people spend their time, the American Time Use Survey and the Multinational Time Use Study[1][3]. It then ran a new, demographically representative survey of 3,995 respondents asking where they would most value help from a personal assistant[1]. The resulting tasks span caregiving, health, education, household administration, personal finances, shopping and government services[1].
02 / 04
What the tasks look like
OpenAI's examples read like a to-do list. Its current models score 93.1% on comparing and shortlisting kindergartens after a move, 98.0% on organising flood-preparation volunteers for four neighbours, 96.3% on setting up a remote way to watch a game with a grandparent, and 90.4% on building a first-week care and meal plan for a parent newly diagnosed with diabetes[1]. A home-search task scores 100%: the agent finds a house in Luzerne County that fits a budget and commute, calculates the transfer tax and adds the deed to a county records cart for the user to review[1].
Across the whole evaluation, coverage of OpenAI's chart reports an average reward of 49.3% for GPT-5.2, released in December 2025, rising to 80.6% for GPT-6 Astra, released in September 2026[2]. OpenAI says it used LifeVal over the past year to track and drive those gains, crediting faster and more capable browser use, computer use and connectors[1].
03 / 04
The caveats
LifeVal is OpenAI grading OpenAI. The announcement doesn't publish the task set, say how many tasks there are or explain how "reward" is scored, and it only shows OpenAI models[1]. Until the tasks are released, outside researchers can't reproduce the numbers or line them up against Claude, Gemini or open models.
04 / 04
Why it matters
Most well-known benchmarks test knowledge, maths or coding. LifeVal is a statement about what OpenAI thinks its more than one billion weekly ChatGPT users actually need: an agent that can finish errands[1]. If the task set is opened up, it could become a useful yardstick for personal agents; for now it is best read as a progress report on where ChatGPT is heading.
FAQ
Frequently asked questions.
01What is LifeVal?
An evaluation OpenAI built to measure how well its models complete practical, everyday tasks end to end, such as finding childcare, accessing government services, budgeting or ordering groceries, rather than just giving advice in chat.
02How were the tasks chosen?
OpenAI started from the American Time Use Survey and the Multinational Time Use Study, then ran its own demographically representative survey of 3,995 people to find where they most wanted help. The tasks cover caregiving, health, education, household administration, personal finances, shopping and government services.
03How do OpenAI's models score?
OpenAI's page shows example tasks scoring between 90.4% and 100%, such as 93.1% for comparing and shortlisting kindergartens after a move. Coverage of OpenAI's chart reports an average reward rising from 49.3% for GPT-5.2 to 80.6% for GPT-6 Astra. These are OpenAI's own figures.
04Can other labs or researchers run LifeVal?
Not as published. The announcement doesn't release the task set, the number of tasks or the grading method, and it only shows OpenAI models, so the results can't yet be independently reproduced or compared across labs.
05Why does this matter?
It signals where OpenAI is steering ChatGPT: toward a personal agent that uses browsers, computers and connectors to get errands done. OpenAI says it has used LifeVal over the past year to track and drive that improvement.
That’s the story
Building with what’s new?
Spectrum Web Co ships products, places talent, builds social narratives and builds custom business tools as monthly services.
Related stories