OpenAI wants to measure AI against everyday life with LifeVal

Childcare, paperwork, finances: OpenAI creates LifeVal to assess its models’ ability to carry out everyday real-world tasks.

Finding a school after a move, organizing volunteers in the face of a flood, managing a family budget, or accompanying a sick parent. All ordinary situations that escape traditional evaluations of artificial intelligence, which are generally focused on mathematics, programming, or scientific knowledge.

With LifeVal, OpenAI proposes measuring something else: the ability of its models to concretely help users in their daily lives. The goal is no longer simply to obtain a relevant answer, but to verify whether an AI can carry out a succession of actions leading to a useful result.

This direction accompanies the gradual transformation of ChatGPT, which the company wants to evolve from a conversational assistant into a tool capable of carrying out procedures, coordinating information, and interacting with different services. 3,995 people surveyed to define priorities

Rather than building its evaluation solely around technical skills, OpenAI started from the activities that actually occupy people’s time.

The researchers relied on two major existing surveys, the American Time Use Survey and the Multinational Time Use Study, which document how populations divide their days between work, household obligations, leisure, and family responsibilities.

OpenAI supplemented this data with a survey of 3,995 people, forming a sample described as demographically representative. Participants were asked, among other things, to identify the activities for which assistance from artificial intelligence would be most valuable to them.

The responses made it possible to define several priority areas: caregiving and support for loved ones, health, education, household administration, personal finances, shopping, and access to public services.

The choice of these categories reveals a shift in the way models are evaluated. An AI can excel in a scientific demonstration while being unable to properly compare schools, organize administrative procedures, or help a family make a decision. From 49.3% to 80.6% in less than a year

OpenAI says it has been using LifeVal for a year to track its models’ progress and guide their training.

The results published by the company show significant progress between GPT-5.2, which obtained an average score of 49.3% in December 2025, and GPT-6 Astra, which reached 80.6% in September 2026.

In between, GPT-5.4 posted 73%, GPT-5.5 reached 77.4%, and GPT-5.6 obtained 80%. These values correspond to an unweighted average of the rewards assigned to the models’ performance on benchmark tasks, and not to an overall success rate for missions actually completed in real-world conditions.

The progress is particularly marked between GPT-5.2 and GPT-5.4, with a gain of 23.7 points. It then slows noticeably: the gap between GPT-5.6 and GPT-6 Astra is only 0.6 point.

This narrowing of scores does not necessarily mean the models are reaching their limits. However, it does underline that gains are becoming less spectacular on this evaluation, despite successive new generations.

OpenAI also highlights improvements in speed and cost. According to its estimates, GPT-6 Astra performs these tasks about 23% faster than GPT-5.6, while the latter achieves broadly similar performance for an API cost estimated to be twice as low. Finding housing, organizing an emergency, helping a loved one

The examples presented by OpenAI give an idea of the situations used to test its models.

In a real-estate scenario, the assistant must search for a house matching a family’s budget, check for the presence of an independent unit, take into account the home-school-work commute, and then prepare an administrative procedure related to the purchase. The best result shown reaches 100%.

Another exercise asks the model to coordinate volunteers to prepare four homes for a possible flood. The model must structure the interventions and produce an action plan adapted to the residents’ needs. The displayed score is 98%.

Family situations also occupy an important place. OpenAI presents, in particular, an exercise focused on enrolling a child in a new preschool after a move, evaluated at 93.1%, as well as a scenario involving dietary and organizational support for a parent recently diagnosed with diabetes, scored at 90.4%.

Another task consists of setting up a simple way to watch a sporting event remotely with one’s grandmother, with a backup plan if the broadcast encounters a problem.

These percentages correspond to the best scores presented for certain exercises. They do not guarantee either systematic success or complete autonomy for the assistant. Several demonstrations end, moreover, with a request for validation or a decision left to the user. The shift from conversation to execution

LifeVal is part of a broader evolution of ChatGPT, which now relies on web browsing, computer use, and connections to external services.

These features change the nature of the tasks accessible to an artificial intelligence. Where a conversational assistant could explain how to carry out a procedure, an agent can now search for information, complete certain steps, prepare documents, or coordinate several successive actions.

This distinction is at the heart of OpenAI’s research publication. For the company, the relevance of a personal assistant should no longer be assessed solely through the quality of its responses, but also according to its ability to produce a usable result.

The question of reliability then becomes central. An error in a general answer can be corrected in a further exchange. An error made during a school enrollment, a financial transaction, or an administrative procedure can have much more concrete consequences.

Autonomy therefore requires verification mechanisms, appropriate authorizations, and the ability to maintain human control over sensitive decisions. A benchmark that still needs to be independently tested

LifeVal brings a different perspective to evaluating artificial intelligence, but its interpretation calls for several reservations.

The published results