OpenAI unveils LifeVal, a test of how well AI handles everyday chores
It is a yardstick OpenAI uses to build ChatGPT, not a tool users can run, and so far it scores only OpenAI's own models.
OpenAI on Monday published LifeVal, a test it uses to measure how well its AI handles everyday jobs such as finding childcare, making a budget or ordering groceries, the company said in a blog post and research paper.
A benchmark is a fixed set of tasks used to score AI models so they can be compared over time. Most well-known AI tests measure school-style knowledge. OpenAI said LifeVal instead measures whether ChatGPT can take actions that get real chores done.
What it means for ChatGPT users
LifeVal is not a feature inside ChatGPT. It is a yardstick OpenAI uses while building its models. OpenAI said that over the past year it has used the test to train models on using a web browser, operating a computer and working with connected apps, part of turning ChatGPT, which it says has more than 1 billion weekly users, into an assistant that can carry out tasks rather than only answer questions.
The paper points to where users may still see trouble. The most common failure was missing information buried in a person's earlier conversations, such as a detail mentioned weeks before, OpenAI wrote.
Can people use it, or compare AIs with it?
Not yet. Anyone can read OpenAI's post and paper, but OpenAI has not released the 169 tasks or code to run them, and neither offers a download. The tasks were written from task writers' own ChatGPT conversations and saved memories, with personal details removed, the paper says.
The paper scored only OpenAI's models, from GPT-5.2 through GPT-6 Astra, so it does not compare ChatGPT with rivals such as Google's Gemini or Anthropic's Claude. Without the tasks, outside companies and researchers cannot run the same test on other AIs.
How the test was built and graded
To pick the tasks, OpenAI started with two big studies of how people spend their time, the government's American Time Use Survey and the Multinational Time Use Study, then surveyed people about where they most wanted help. The tasks cover caregiving, health, school, household paperwork, money, shopping and government services, and run in simulated settings with a saved copy of the web.
Each task has a checklist of what a good answer must include, and an AI grader built on GPT-5.6 scores the work. On a 30-task check, human graders and the AI grader gave similar average scores, about 84%, OpenAI said. The company built the test, graded its own models and did not mention any outside review. The paper notes limits: the time-use data ends in 2021, its survey covered only U.S. English speakers, and the saved web can vary between runs.
OpenAI posted sample scores for its current models: 93.1% on shortlisting kindergartens after a move, 98% on organizing flood volunteers, 100% on finding a home that fits a budget and 90.4% on a meal and care plan for a parent with new diabetes.
Source: OpenAI, "LifeVal," Nov. 2, 2026 (openai.com), and OpenAI research paper "LifeVal" (cdn.openai.com)