OpenAI has published LifeVal, a benchmark, meaning a fixed set of tasks used to score an AI system, that measures how well its models handle ordinary life admin: finding childcare, building a budget, dealing with government services, and shopping. The company describes it on its own site, in a post dated 2 November 2026, and says it has used the test over the past year to track and steer progress.
The task list did not come from engineers brainstorming. OpenAI started from two large time-use studies, the American Time Use Survey, plus, for other countries, the Multinational Time Use Study, which log what people actually do with their days. It then ran a new survey of 3,995 respondents, described as demographically representative, asking which of those tasks people would most want a personal assistant to take on. The task areas that came out of it run from caregiving and health to schooling, household paperwork, money, shopping, and dealing with government offices.
OpenAI showcases five example tasks with scores, and no overall result. Enrolling a child in the right kindergarten after a move scored 93.1 percent. Organizing volunteers ahead of a flood scored 98.0 percent. Setting up a way to watch a game remotely with a grandmother scored 96.3 percent, and a first-week meal plan for a parent facing a new diabetes diagnosis scored 90.4 percent. Finding a family home in Luzerne County within a budget and a short school run scored 100 percent, with the model locating one house, calculating the transfer tax, and adding the deed to a county records cart for review.
Those five numbers invite a question the post does not answer. OpenAI does not say which model produced them, how a task is graded, or what a score of 93.1 percent means in practice. The post gives no overall LifeVal score, only a line that the “latest models score very highly on many LifeVal tasks.” A chart of performance over time is mentioned, but the model names and the scale behind it are not spelled out in the text.
The conflict of interest is plain. OpenAI chose the tasks, wrote the prompts, and grades the results, and no outside party has reproduced any of it. That is normal for a company’s internal progress tracker, and it is also why the scores work better as a statement of direction than as a comparison. A test whose showcased examples land between 90 and 100 percent also leaves little room to show further gains, unless the full set contains much harder tasks than the ones OpenAI chose to publish.
The direction itself is clear. OpenAI says its models have become faster, more capable, and better aligned when driving a browser, operating a computer, and using connectors (the integrations that let ChatGPT reach other apps and services). It frames this as moving ChatGPT beyond a chat tool toward an agentic personal assistant, one that takes actions rather than only giving advice. With more than a billion weekly ChatGPT users, by OpenAI’s count, the everyday-task market is where the company wants its agents judged, and a survey-built benchmark gives it a way to say so with numbers.
Most of the example tasks end with the model handing a decision back to the person: two kindergarten preferences left for the user to call, a diabetes meal plan offered for review before sharing, a deed added to a cart rather than filed. That pattern is the useful part for anyone building consumer or government-facing agents. It suggests the version OpenAI considers a success stops short of acting alone on anything with consequences.
Builders who want to know whether these results hold should wait for OpenAI to publish the full task set and grading method, or run their own versions of the same chores against the models they use.
Reported by OpenAI, in its LifeVal research post of 2 November 2026.
