Press "Enter" to skip to content

Prince Mario-Max Schaumburg-Lippe: OpenAI LifeVal Tests AI on Everyday Tasks

OpenAI has released LifeVal, a benchmark that asks a simpler question than most AI tests: can an assistant handle the jobs people actually hand it during an ordinary week? The release was reported on November 2 and 3, 2026. No exam hall. No puzzle sheet. Just the paperwork, coordination and small decisions that eat up an afternoon.

That shift sounds modest. It is not. For years, the public scoreboard for AI has leaned on tests that look like school: maths problems, coding challenges, knowledge questions with one right answer. Useful, up to a point. But most people do not hire help to sit an exam. They want the form filled in, the appointment moved, the budget checked, and the right document sent to the right person.

Built from how people spend their time

LifeVal’s tasks were built from time-use surveys and from feedback in a poll of 3,995 people. That grounding matters. Instead of inventing clever puzzles, the designers started with where hours actually go, then turned those activities into tasks an assistant can be scored on.

The coverage is broad and deliberately everyday. The list includes personal administration, caregiving coordination, food assistance paperwork, healthcare navigation and personal finance. There are 16 primary task categories in total. Read that list again and you can picture real inboxes: a letter to chase, a relative’s appointment to arrange, a benefit form that needs the right evidence attached.

Why everyday tasks are hard to score

Everyday work is messy in a way exam questions are not. A form may need information from three places. A good answer may depend on a deadline buried in a letter. Coordinating care for someone else involves other people’s schedules, preferences and privacy. Half the job is knowing what done looks like.

That is why a benchmark in this area has value even before you look at any score. It forces a shared definition of success. Did the assistant complete the paperwork, or just draft a nice summary of it? Did it book the thing, or talk about booking the thing? Anyone who has tested a new assistant, human or software, knows that gap well.

We saw a similar move toward real work in coding agents judged on finished pull requests. The pattern is the same across fields. Show me the completed task, not just the fluent reply.

The headline result, with the label it needs

OpenAI reports that GPT-6 Astra scored an 80.6 percent overall average on LifeVal, up from 49.3 percent for GPT-5.2 nine months earlier. OpenAI also reports 100 percent completion rates in 10 of the 16 categories. Be clear about the source: these are OpenAI’s own reported results on its own benchmark. That does not make them useless. It does mean you should read them as a progress claim from the lab that built both the test and the model, and look for outside checks over time.

Even with that label, the size of the reported jump deserves attention. A move from about half of everyday tasks completed to about four in five, in nine months, suggests the gains are landing where users feel them. Not in a trivia score. In whether the assistant finishes the job you gave it before dinner.

Completion in 10 of 16 categories at 100 percent also tells a useful story about uneven progress. Some kinds of admin are now, on these reported results, effectively solved in test conditions. Others still lag. That unevenness is honest, and it matches experience. Assistants tend to be brilliant at one chore and oddly shaky at the next.

What this changes for buyers and builders

For anyone choosing an assistant for a team or a household, LifeVal points to a better set of questions. Do not start with, which model is smartest? Start with, which of our weekly tasks does it finish without rescue? Bring your own list: the five recurring jobs that consume the most time. Test those. Time them. Count the corrections.

For builders, the lesson cuts the other way. If measurement moves toward lived tasks, products need to be designed for completion, not just conversation. That means clear confirmations, a record of what was sent where, and an easy way to undo a step. A cheerful paragraph is not a booking.

The same practical standard appears in software that lets builders query their systems in plain language. The promise only holds if the answer leads to a correct, finished action in the real workflow.

Practical takeaways you can use this week

Keep a small scorecard of your own. Pick 10 tasks you delegate often. Note whether the assistant finished each one, needed one prompt of help, or needed you to take over. After a month, you will have a benchmark that matters more to you than any launch chart, because it is drawn from your life.

Give tasks in a way that helps completion. State the end state you want, the deadline, and what proof of done looks like: a confirmation number, a sent email you are copied on, a file in a named folder. Assistants, like people, do better work when done is defined up front.

And keep sensitive steps under your own hand. Healthcare navigation and personal finance sit inside LifeVal’s scope, and they are exactly the areas where a person should review before anything is submitted or paid. Speed is welcome. A final human check is still cheap insurance.

LifeVal will not settle every argument about how good AI assistants have become. No single benchmark can. But it points the argument at the right place: the kitchen table, the inbox, and the pile of forms by the door. If assistants keep improving there, most people will not need a chart to notice.


Also published on:

More from AI Artificial IntelligenceMore posts in AI Artificial Intelligence »
More from BREAKING NEWSMore posts in BREAKING NEWS »

Be First to Comment

Leave a Reply

Your email address will not be published. Required fields are marked *