LifeVal tests everyday AI agents: they scored 90.4–100% across five tasks

Most AI benchmarks are designed for laboratory convenience: a model is given a question, then its answer is compared with a reference answer. That is not enough for an everyday agent. It may correctly explain how to choose a daycare center while failing to account for travel time, the child's age, application deadlines, and the documents required by a particular family. Users judge not how elegant the advice sounds, but whether it brings the task closer to completion.

OpenAI introduced LifeVal, a suite of evaluations for these kinds of multistep everyday tasks. In the published examples, models scored between 90.4% and 100%: 90.4% for a nutrition and care plan following a diabetes diagnosis, 93.1% for selecting a daycare center, 96.3% for organizing a group match viewing, and 98% for coordinating volunteers ahead of a flood. The housing-search task received a score of 100%. Importantly, these are individual demonstration tasks, not LifeVal's average accuracy.

What LifeVal actually evaluates

The authors began not with a list of model capabilities, but with a question: which activities do people genuinely need help with? To answer it, they combined time-use research from the United States and other countries with their own survey. OpenAI's project page reports 3,995 participants, while the paper refers to more than 3,800 respondents. The resulting map includes 148 categories, including caring for loved ones, education, personal finance, shopping, moving, job searches, and access to government services.

The researchers then created personalized tasks for high-demand scenarios. These incorporate a person's context, reference files, memories, and anonymized excerpts from real ChatGPT conversations. Task creators had to have personal experience solving the relevant problem, write the evaluation criteria, and pass at least three stages of quality review.

Consider a simple example: a family is moving and asks for help choosing a daycare center. A good agent does not stop at listing facilities. It compares distance, admission requirements, and the family's preferences; gathers suitable options; breaks the enrollment process into steps; and separately flags decisions that cannot be made without a person. The entire result is evaluated, not one successful recommendation.

Why this matters to developers

LifeVal shifts the unit of measurement from an answer to a completed scenario. For a product team, that provides a practical target: an agent test should include search, work with browsers and files, use of personal context, a request for confirmation before a sensitive action, and a clear outcome for the user.

This approach helps uncover errors that remain invisible in an ordinary chat. An agent may find the correct form but select the version for another region. It may prepare a budget but omit a recurring payment. It may confidently continue working despite lacking one critical condition. It is therefore useful to measure not only the quality of the final text, but also the share of completed steps, the number of unverified assumptions, the accuracy of sources, and the ability to hand a decision back to a person at the right time.

Where caution is required

LifeVal was created by OpenAI and has already been used to improve its models. It is a useful internal tool, but the published high scores cannot automatically be generalized to every prompt, country, or interface. The page provides no single aggregate score that would allow all models to be compared fairly, while personalized scenarios are especially sensitive to the completeness of context and the tools available.

There is also a product risk: an everyday agent works with finances, health, family circumstances, and government documents. The more realistic the evaluation, the stricter the rules must be for data anonymization, access control, action logging, and transaction confirmation.

The practical conclusion is straightforward: teams should build their own small LifeVal with 20–30 real user scenarios. For each one, they should define the initial data, permitted actions, mandatory checks, the point at which control must be handed to a person, and a measurable outcome. Such a suite will reveal an agent's product readiness faster than yet another general-knowledge benchmark.