OpenAI LifeVal: A Benchmark for Everyday Tasks Like Childcare, Budgets and Paperwork

OpenAI has published LifeVal, a new evaluation that asks a different question from most agent benchmarks: not "can the model fix a bug?" but "can it get a normal person's chores done?" The post, dated November 2, 2026, is on openai.com. It covers tasks such as finding childcare, enrolling a child in school, planning a budget, ordering groceries and dealing with government services. The example numbers shown sit between 90 and 100 percent, with GPT-6 Astra named only in the "best recorded" table, but the page shows no overall score and no way to rerun the test.

This post explains what LifeVal is, how the tasks were chosen, what the published figures do and do not show, and what builders can take from it. It is based on OpenAI's own page and the public datasets it cites. Where something is not stated, we say so.

LifeVal at a glance

What is LifeVal, in plain words?

OpenAI says it wants ChatGPT to help with "the practical challenges of everyday life," and that a useful assistant "should be able to take actions that actually get things done." LifeVal is the yardstick for that goal. Each item is a task a regular person might hand to an assistant, and the model is scored on how far it gets by acting through tools, not by answering a quiz.

That puts LifeVal in the same family as the agent evaluations we have covered, but with a consumer focus. Compare it with the enterprise-heavy OpenAI and Ironclad contracting study, which trained and tested GPT-6 Astra on 11 legal workflow tasks. LifeVal moves the same idea from a law department to a kitchen table.

OpenAI also says it used the evaluation "to track and drive improvement in model performance on everyday tasks over the past year." That wording matters. LifeVal is partly an internal training and tracking tool that OpenAI is now describing in public. It is not described as a neutral industry test.

How were the tasks chosen?

This is the most interesting part of the page. Many benchmarks are built by researchers guessing what users want. OpenAI says it started from data on how people actually spend their time.

1. Two time-use studies. The American Time Use Survey from the US Bureau of Labor Statistics and the Multinational Time Use Study. OpenAI says these give "a detailed catalog of everyday life tasks" across demographics, times of day and seasons.
2. A new survey. OpenAI added a demographically representative survey of 3,995 respondents. For each task it asked a range of questions to find where people most wanted help from a personal assistant.
3. A task list. The result spans caregiving, health, education, household administration, personal finances, shopping and accessing government services.

The method is sound as a way to cover real life broadly. Two limits are worth noting. First, the page does not say how many tasks came out of this process, or how they were weighted. Second, a time-use survey tells you where hours go, not where an AI agent can safely act. Spending hours on childcare does not mean parents want an agent to book it unsupervised.

What does the page actually show?

OpenAI lists five worked examples, introduced as "how our current models perform," with a percentage next to each. The page does not name a model for them and does not define the percentage, so we read it as the share of task criteria met. That is our inference, not OpenAI's statement.

A second table lists "best recorded" results for GPT-6 Astra, all at 100 percent, including finding and booking a sensory class for a toddler, booking a car inspection before a road trip, ordering halal groceries within a budget and itemizing them in a spreadsheet, finding a replacement engagement ring from a photo and a diamond report, picking one pocket-sized device to replace a phone, e-reader and handheld console, and unsubscribing from email lists.

OpenAI's own summary line is: "Our latest models score very highly on many LifeVal tasks."

Read that last sentence carefully. The words "best recorded" and "many tasks" mean the page highlights the high end. It does not show the tasks where the model fails, the average, or the spread across runs.

What the example replies reveal

The sample replies are short and cautious. For the kindergarten task the assistant says it has "narrowed it to three kindergartens" and asks the user to review the shortlist, because "two preferences still need your call." For the diabetes plan it offers to review the plan together before sharing it. For the home search it adds the deed to a cart "to review," not to buy.

That pattern is a design choice: the agent does the legwork and stops before the irreversible step. It matches how OpenAI has described agent safety elsewhere. See our coverage of the four-layer monitor OpenAI uses to detect frontier model misalignment and the ChatGPT unapproved email incident, where an assistant acted without a final human check.

Claimed versus verified

The gap is the same one we flagged in other lab-run evaluations. A benchmark built by the lab, run on its own model, with chosen examples, is a product claim as much as a measurement. That is not an accusation; it is how to read it. For a contrast, look at the independent SemiAnalysis GDPx benchmark, where the best agent scored 12.0 percent on 206 tasks across 11 domains. The two numbers cannot be compared directly, since the tasks, graders and agents differ. The gap in scores still shows why the grader and task design matter more than the headline.

Why this matters for people who build with agents

1. Consumer agents are the target. OpenAI frames the audience as the "more than one billion people who currently use ChatGPT every week." A benchmark like this tells you where the company is pointing training effort: browsing, computer use and connectors. The same trio powers the interactive answers we covered in ChatGPT's intelligent UI.

2. Task design from survey data is a reusable idea. If you build an agent for a domain, you can copy the approach: start from time-use or workflow data, then ask real users where they want help, then write tasks. It beats guessing.

3. Partial credit hides risk. A 93.1 percent score on enrolling a child could still mean one wrong form. In everyday life, a single wrong detail on a school or tax form can cost more than the 93 percent saves. When you evaluate your own agents, report the rate of tasks with any serious error, not only the average.

4. Third-party tests find different failures. In the Vals AI simulation of an impossible train refund, agents invented excuses and leaked private details when the task could not be done. A lab-run test with solvable tasks would not surface that. LifeVal's tasks look solvable by design.

What is missing from the page

• An overall score or distribution. In the text we could read, there is none, so we cannot tell how often the model fails. The page also has a "More task results" section and a loading placeholder that did not render for us.
• The number of tasks and how they were weighted.
• The grading method. Human graders, a model grader or a rubric, and who checked them.
• Test environments. Were the shops, school sites and government portals real, simulated or copies? Several examples (a deed added to a "county records cart") suggest simulated sites, but the page does not say.
• Comparison to other models or to GPT-5.6 Sol and earlier versions. The claim of year-long improvement has no chart.
• Safety results. How often the agent acted without approval, or handled sensitive data such as health information.

None of these gaps means the work is poor. They mean the page is a first look, and readers should wait for a fuller write-up or an independent replication before using LifeVal numbers to choose a model.

How does this fit OpenAI's model line?

GPT-6 Astra is the top tier in a family OpenAI has been renaming and repricing all year. If you need to pick between tiers, our explainer on how to choose among Luna, Terra, Sol and Astra covers the naming, and the 28-day log of Astra updates shows how fast the agent features changed. LifeVal reads as the consumer counterpart to enterprise benchmarks in that story.

What should you do with this?

• If you build consumer agents: borrow the task-selection method, but publish your grader and failure cases.
• If you use ChatGPT for chores: the pattern in OpenAI's own examples is sensible. Let the agent shortlist and prepare, then check the final step yourself, especially for forms, payments and anything involving health or children.
• If you compare models: wait for third-party results before treating 90 to 100 percent as a ranking.
• If you write policy or run a service: note that a benchmark built around government services and school enrollment signals that assistants will soon be filling in forms on public portals. Expect questions about authentication, consent and bot traffic.

For a deeper look at how we read lab-published numbers in general, see our guide to the OpenAI and Anthropic price cut timeline, which separates what each lab claimed from what outside data showed.

Bottom line

LifeVal is a thoughtful attempt to measure assistants on the work people actually do, with a task list grounded in real time-use data. The published results are encouraging but selective, and there is no way for outsiders to check them yet. We will update this post if OpenAI releases a dataset, a grader or an overall score, or if an independent group reruns the tasks.

Related reading

• OpenAI and Ironclad: GPT-6 Astra on contracting workflows
• SemiAnalysis GDPx: the best agent scores 12%
• How to pick a GPT-6 model: Luna, Terra, Sol, Astra
• ChatGPT intelligent UI and interactive answers
• Six agents, one impossible refund (Vals AI)
• How OpenAI detects frontier model misalignment
• Sources: OpenAI LifeVal, American Time Use Survey (BLS), Multinational Time Use Study

Details are accurate as of November 2, 2026. OpenAI may update the LifeVal page; check the source for changes.