OpenAI's LifeVal Tests AI on Everyday Errands, and GPT-6 Astra Scores 80%

OpenAI has published LifeVal, an internal test that scores how well its AI models handle everyday chores such as picking a kindergarten, planning meals for a relative with diabetes and filling a grocery cart. The company says its GPT-6 Astra model, released in September, scored 80% overall on the evaluation, up from 49.3% for GPT-5.2, a model from December 2025.

The announcement, dated November 2, 2026, matters because it shows where OpenAI is aiming its training effort next: not exam questions or coding puzzles, but the errands and paperwork that fill ordinary weeks. It also arrives with a big caveat. LifeVal is run by OpenAI, on OpenAI's models, and outsiders cannot yet check the numbers.

Key facts at a glance

• What: LifeVal, an OpenAI evaluation of AI performance on practical, everyday tasks, published on November 2, 2026.
• Headline result: OpenAI says GPT-6 Astra scored 80% overall, against 49.3% for GPT-5.2.
• How tasks were chosen: two large time-use studies plus a new survey of 3,995 respondents.
• What it covers: caregiving, health, education, household administration, personal finances, shopping and government services.
• Access: no public dataset, grading code or leaderboard has been released.

Why OpenAI built an everyday-life test

Most well-known AI benchmarks measure things with a clear right answer: a maths result, a passing unit test, a correct fact. Daily life rarely works that way. Choosing a school depends on an address, a budget, opening hours, commute times and a child's needs, and two families can reasonably want different answers.

OpenAI frames LifeVal around its user base, saying more than one billion people use ChatGPT every week. In its announcement the company says a truly helpful assistant "should be able to take actions that actually get things done", rather than only offering advice. It adds that over the past year it trained its models to become faster and more capable at browser use, computer use and connectors, the features that let ChatGPT open websites, operate software and reach a user's other apps.

LifeVal also has a sibling. OpenAI previously released GDPval, which measures AI on economically valuable work tasks across industries. LifeVal moves the same idea from the office to the kitchen table, and OpenAI says it has already used the test internally for about a year to track and drive model improvements.

How the tasks were chosen

OpenAI says it started with two long-running studies of how people spend their time: the American Time Use Survey, run by the US Bureau of Labor Statistics, and the Multinational Time Use Study. Together they catalogue everyday activities across demographics, times of day and seasons.

That raw list had an obvious problem. Some of the biggest time sinks, such as sleeping and eating, are not things an AI assistant can do for anyone. So OpenAI ran its own demographically representative survey of 3,995 respondents, asking where people most wanted help from a personal assistant.

OpenAI researcher Jenny Yang said a smaller group of about 100 users then wrote down tasks they wanted done, the personal context they would give the AI, and a rubric for judging whether the job was done well. Contributors had to show real familiarity with their topic. For a financial planning task, for example, the contributor was expected to have worked with a financial adviser for several years.

"You had to like have evidence that you were some kind of like expert or really understood what the right answer looked like for you," Yang said.

The scoring is rubric-based rather than pass or fail. A school search can be judged on tuition, distance, schedule and eligibility, and a grocery order on dietary needs, price and quantity. According to OpenAI, the test looks at whether a model can navigate websites, pull relevant details from a user's chat history and reach conclusions that fit that person's needs.

What the published scores show

OpenAI's page highlights five worked examples, each with a percentage score. The page credits them to "our current models", and the company has attributed these results to GPT-6 Astra.

Source: OpenAI's LifeVal announcement, November 2, 2026.

A second list of "best recorded" GPT-6 Astra results shows 100% on tasks including booking a sensory class for a toddler, booking a car inspection before a road trip and ordering halal groceries for three favourite meals. OpenAI's own summary is that its latest models "score very highly on many LifeVal tasks".

The sample replies are notably cautious. In the kindergarten task the assistant narrows the list to three options and says two preferences "still need your call". The home search ends with the deed added to a cart "to review", not purchased. The diabetes plan is offered for joint review before it is shared. In each case the model does the legwork and stops before an irreversible step.

The competitive backdrop

LifeVal lands in the middle of a race to build AI agents that act on a user's behalf. Meta's Muse agent, launched in September, reached 10 million US downloads in its first 40 days, according to market intelligence firm Sensor Tower. OpenAI introduced its own Dots agent, which it says can work on long-running tasks in the background, the same month, and Anthropic announced its Claude Duet agent app in October.

A benchmark that rewards finishing real errands gives OpenAI a way to show progress on exactly the ground its rivals are contesting. Yang described the goal as giving people time back, whether by cutting out chores such as returning a package or by helping them learn new skills.

The other side: what outsiders cannot check

Independent AI analysts were quick to point out the limits. LifeVal is an OpenAI-controlled evaluation, not a reproducible public benchmark. Outside researchers cannot inspect the full task set, the hidden test split, how rubrics are weighted, how graders were calibrated, the browsing environment or how many attempts each model got.

Several practical gaps stand out:

• No full distribution. The page shows selected high scores and a "best recorded" list, not the tasks where models struggled.
• No rival models. Only OpenAI models appear, so the numbers cannot be compared with Anthropic, Google or Meta systems.
• Partial credit can hide errors. A 93.1% score on a school search could still include one wrong detail on a form, which in real life might matter more than everything the agent got right.
• Tuning versus testing. OpenAI says it used LifeVal to drive improvements for a year, which raises questions about how cleanly the test tasks are kept separate from training.

There is a contrast within the same day's news. A separate, independent benchmark of expert-level professional tasks launched on November 2 put the best frontier model at just 12 out of 100. The tests are very different and the numbers are not directly comparable, but the gap shows how much task design and grading drive the headline figure.

What it means for you

For everyday ChatGPT users in the US, UK, Canada and Australia, LifeVal is a signal that agent features for shopping, bookings and forms will keep expanding. OpenAI's own examples model a sensible habit: let the assistant research and prepare, then check the final step yourself, especially for payments, government forms, health plans and anything involving children.

For developers building agents, the method is the useful part. Starting from time-use data, asking real users where they want help and writing scenario-specific rubrics is a pattern any team can copy without access to OpenAI's private task set. Recording the model version, tools, retries and settings alongside every score makes results far easier to trust.

For buyers and IT teams comparing models, treat LifeVal scores as a vendor claim until independent results appear. A self-reported 80% does not tell you how a model will behave on your own workflows or data.

What happens next

OpenAI has not said whether it will publish the LifeVal tasks, grader or an overall score breakdown. The open questions are clear:

• Will outside groups be able to rerun LifeVal, or a public subset of it?
• Will future OpenAI model launches report LifeVal results alongside other benchmarks?
• Will rivals publish comparable everyday-task evaluations for their own agents?
• How will safety results, such as how often an agent acts without approval, be reported?

With GPT-6 Astra already the reference point, the next OpenAI model release is the natural moment to see whether LifeVal becomes a regular part of how the company reports progress.

Conclusion

LifeVal is a serious attempt to measure AI on the work people actually do at home, built from real survey data rather than researcher guesswork. Its published results are strong but selective, and nobody outside OpenAI can verify them yet.

Practical takeaways:

1. Keep the final click. When an AI assistant handles bookings, orders or forms, review the result before anything is paid for or submitted.
2. Read scores in context. A self-run, best-case benchmark is a product claim until an independent test confirms it.
3. Borrow the method. If you build or evaluate agents, use real user tasks with explicit constraints and outcome-based rubrics.
4. Log your settings. Record model version, tools and retries with every evaluation so results can be compared over time.
5. Watch for replication. Wait for third-party results before choosing an assistant based on LifeVal numbers alone.

FAQ

Can I run LifeVal myself?

No. OpenAI has not released the task set, grading code or a public leaderboard.

Does LifeVal change ChatGPT or the API?

No new feature, endpoint or pricing change was announced with it. It is a measurement tool OpenAI uses to track model progress.

How is it different from GDPval?

GDPval measures professional work across industries. LifeVal focuses on personal tasks such as childcare, budgets, shopping and government services.