OpenAI launches LifeVal, a benchmark where an agent enrolls children in daycare and orders groceries

Most benchmarks ask a model what it knows. OpenAI is proposing a different question: can it handle the tedious routines of everyday life for you? That is the idea behind LifeVal, an evaluation in which an agent must do more than offer advice—it must complete the task.

What LifeVal is

The project is based on two major datasets about how people spend their time: the American Time Use Survey and the Multinational Time Use Study. OpenAI used them to compile a catalog of everyday activities. It then conducted its own demographically representative survey of 3,995 respondents to identify where people most need a personal assistant.

The resulting tasks cover caregiving, health, education, household administration, personal finance, shopping, and government services. These are not multiple-choice questions. They are scenarios that require opening websites, comparing options, doing calculations, and completing transactions or forms.

What the tasks look like

One example involves a user who has moved and asks the agent to compare daycare centers near the new home. The agent narrows the options to three, gathers enrollment requirements, and clearly states that two preferences require a human decision. The score for this task is 93.1%.

Other examples from the publication include:

• finding a home in Luzerne County within budget, with a separate living unit and a short commute after school, calculating the transfer tax, and placing the document in the cart: 100%;
• preparing the first week of meals and reminders for a parent newly diagnosed with diabetes: 90.4%;
• organizing volunteers ahead of a flood: 98.0%;
• setting up a shared match-viewing session with a grandmother, including a fallback in case the stream fails: 96.3%.

GPT-6 Astra's top-results table shows 100% on tasks such as booking a sensory class for a toddler, scheduling a vehicle inspection, ordering halal groceries within a budget, and unsubscribing from mailing lists.

OpenAI emphasizes that over the past year, models have become significantly faster and more accurate at working with browsers, computers, and connectors. This is what turns ChatGPT from a chat interface into an agent.

What to keep in mind

First, the company created the benchmark, selected the examples, and is presenting the results itself. The text does not describe the scoring method: who decides, and how, that a task was completed at 93.1% rather than 100%. Nor does it compare the results with competing models.

Second, 100% on a handful of tasks says more about saturation on those particular scenarios than about an agent's reliability in real life. Tasks involving payments, healthcare, and government services require a very low error rate, which an average percentage does not demonstrate.

Third, in the examples, the agent often stops and asks the user. That is a good pattern, but it also means end-to-end autonomy has not yet been achieved.

What to do with it

For product teams: treat LifeVal as a map of the scenarios OpenAI intends to pursue. Daycare, budgets, government services, shopping, and caregiving are where agentic ChatGPT is headed, and narrow wrappers around these tasks will come under pressure.

For agent developers: adopt the approach. Build task sets from real user behavior rather than academic tests, and measure whether an action was completed rather than the quality of the generated text. Add human confirmation points wherever errors are costly.

For founders: do not mistake scores of 90–100% for market readiness. Wait for independent verification before building a product on the assumption that the agent can make a purchase or complete a form without errors.