AI takes on the chores: OpenAI's LifeVal puts GPT-6 Astra at 80.6% on everyday tasks
OpenAI's LifeVal tests AI on everyday errands instead of exam questions. GPT-6 Astra averages 80.6%, up from 49.3% for GPT-5.2 — but OpenAI wrote the test, graded it, and only entered its own models.

In a single week, OpenAI published two report cards. One tests whether an AI can work as a lab technician. The other tests whether it can help out around the house. The second one is called LifeVal, published on November 2. They measure different things: the first is about professional judgment, the second about the chores nobody wants to do but nobody can afford to get wrong — finding childcare, enrolling a child in kindergarten, ordering groceries, dealing with government paperwork, planning a week of meals for a parent with a new diabetes diagnosis. OpenAI's bet is that the next frontier for AI is not the exam hall, but all of that unglamorous errand-running.

1. Not what it knows, but what it gets done
LifeVal is built on a different question from most benchmarks. Traditional benchmarks ask what a model knows. LifeVal asks what a model can finish. OpenAI puts it plainly on the official page: when someone wants help making life easier, they need more than advice or a chat response — to be a truly helpful personal assistant, ChatGPT has to be able to take actions that actually get things done. So a task does not end with a good-looking answer. It ends with a result: the kindergartens compared, the transfer tax calculated, the deed sitting in the county records cart waiting for review.
There is one detail in the five worked examples worth singling out. In the kindergarten task, the model narrows the choice to three schools, organizes the enrollment details, and then says two preferences still need the user's call, and asks whether to review the shortlist together. It does not decide for the user, and it does not throw the question back untouched. It does the work up to exactly one decision short of done, and stops there. That is the hardest and most easily overlooked part of everyday tasks: knowing when to stop and when to ask. A good everyday assistant does not decide everything for you — it does 90% of the drudgery and lays the 10% that genuinely needs your judgment out clearly in front of you. LifeVal rewards that kind of restraint with high scores, and that scoring instinct is itself worth more than the scores.
Note also how the tasks are done. LifeVal does not ask the model to sit and write essays. It asks it to use browser use, computer use, and connectors — to browse websites, fill in forms, compare prices, and pull things together. In other words, it tests exactly the agent capabilities that have been talked about all year: the same skills that run tests and edit files when writing code are now pointed at everyday errands. OpenAI says it has used this evaluation internally for the past year to track and drive improvement on everyday tasks — LifeVal was a yardstick inside the company before it became a public announcement.
2. 3,995 people first answered a question: which chores do you most want to offload?
The tasks were not invented by researchers in a meeting room, and that is worth saying up front. OpenAI started by looking at people, then writing the test. Two time-use studies provide the raw material for how a day is actually spent: the American Time Use Survey from the US Bureau of Labor Statistics, and the Multinational Time Use Study, which standardizes more than a million diary days from over a hundred nationally representative surveys. Together they answer a simple question: where does an ordinary person's time actually go?
But "spends time on" and "wants to offload" are two different things. Many people spend hours with their children and have no wish to outsource that; meanwhile, some tasks that take only a small slice of the day — wrestling with a government website, comparing three kindergartens, working out a household budget — cause the most headaches. So OpenAI added a new survey of 3,995 demographically representative respondents, asking item by item where they would most want a personal assistant's help. The seven areas are what survived that filter: caregiving, health, education, household administration, personal finances, shopping, and accessing government services. The examples on the page are concrete: finding childcare, accessing government services, creating a budget, preparing for school, modifying recipes, ordering groceries, designing workout plans. Not one task is "write a sonnet" or "explain quantum entanglement." Every one is the kind of thing that is not hard, but is annoying — and cannot be done wrong.
The method has a blind spot that should be stated plainly. The survey of 3,995 people is described as demographically representative, but OpenAI does not say representative of which population — American users, or users worldwide? The time-use studies themselves are centered on the US and the countries in the multinational study. A commuter in Tokyo and a parent in Texas do not run the same daily errands. LifeVal measures something closer to "a largely US-centered sample of everyday life" than "everyday life for all of humanity." That does not make it worthless, but readers should know the scope of the test when they read the scores.
There is also a background number: OpenAI says more than one billion people now use ChatGPT every week. That is why LifeVal exists. When your user base is ten times a city and several times a country, "helping everyone live their lives better" stops being a slogan and becomes a product line — and LifeVal is the acceptance test for that product line. The question it answers is not how smart the model is, but how much of a billion people's chore lists it can take over.
3. From 49.3% to 80.6%: real progress, but read the denominator first
The official page itself carries no overall score table, only the line "our latest models score very highly on many LifeVal tasks." The full score curve comes from a chart OpenAI published with the announcement, as reported by independent coverage: GPT-5.2 (December 2025) at 49.3% average reward, GPT-5.4 at 73.0%, GPT-5.5 at 77.4%, GPT-5.6 at 80.0%, and the newest GPT-6 Astra at 80.6%. In under a year, the line climbed from below passing grade to above 80%. That is genuine progress. And it is not only the score that improved: Astra is reportedly about 23% faster than GPT-5.6 at these tasks, while GPT-5.6 delivers comparable performance at roughly half the estimated API cost. So the year's gains are not only in "can it do it" but in "how fast" and "how much it costs."
But the number 80.6% can only be read with its definition attached. It is an "average reward," not a task completion rate. Each task earns partial credit for how far it gets, and the task averages are combined with equal weight. Enrolling a child in kindergarten can score above 90% if the shortlist is right, the costs are worked out, and only two preferences are left for the user to decide — but that is not the same as "eight times out of ten, you never have to get involved." OpenAI also lists six tasks where Astra's "best recorded" score is 100% — booking a car inspection before a road trip, ordering halal groceries within a budget and itemizing the order in a spreadsheet, unsubscribing from email lists. Best recorded, note, not average. The distance between a best attempt and a typical attempt is exactly the part that frustrates people in daily use: pulling off a task beautifully once is not hard; doing it reliably every time is.
So what the curve really shows is agents moving from "often breaks down" to "mostly gets it done." Anyone who works with AI daily can feel that shift: a year ago, asking a model to work through a government website would usually stall halfway; now it mostly finishes the process and hands only the final decisions back to you. The progress is real, but 80% is still some distance from "safe to hand over."
The curve also has a practical message for anyone choosing a model: the most expensive option is not necessarily the answer. GPT-5.6 scores 80.0%, only 0.6 points below Astra's 80.6%, at reportedly about half the estimated API cost; Astra is about 23% faster. If you run everyday tasks in bulk — comparing prices, filling forms, organizing material for dozens of users a day — 5.6 is clearly the better value; if you need a single task done as fast as possible and cannot wait, Astra is worth the premium. A benchmark's value is not only in the ranking but in this kind of "good enough" arithmetic. The pity is that, for now, the arithmetic can only be done inside OpenAI's own product line, because no rival model has been run on the same test.
4. The biggest problem with the test: the same company wrote it, graded it, and took it
LifeVal's weakness is not in how the tasks were chosen. It is in who chose them, who graded them, and who sat the test. The answer is the same company in all three cases: OpenAI designed the tasks, set the grading rubric, and ran the tests, and only its own models appear in the results. The official page publishes no full task set, no grading rubric, and no dataset or code for others to rerun. No rival lab's model appears in any table. CNN's coverage calls LifeVal "a new internal tool," which is an accurate description: first it is a yardstick OpenAI uses to measure its own progress, and only second a public scorecard.
That is not an accusation of faking the numbers. It is a limit on what the numbers can be used for. A ruler that only measures yourself can tell you "how much better we are than last year," but it cannot answer "how do we compare with anyone else." More importantly, the choice of tasks can itself lean toward what the test-maker is good at. OpenAI says on the page that the tasks drew on real requests from about 100 users, according to CNN's reporting, with participants supplying real circumstances and helping set the criteria for a satisfactory result. That is more solid than inventing tasks from scratch, but the sample still comes from OpenAI's own user pool, and its coverage is limited by definition.
One more detail is worth noting: the five worked examples score suspiciously neatly, between 93.1% and 100%. Examples are chosen to show the model off, so naturally they are tasks where it performs well. How large the full task set is, what share of it is hard, and how the models do on the hardest slice — the page says nothing about any of that. OpenAI researcher Jenny Yang told CNN that LifeVal is meant to help the company identify both improvements and remaining weaknesses. That is a fair statement, but where exactly the weaknesses are is not something the public can see yet.
So the right way to read this scorecard is as OpenAI's internal progress report, not as an industry leaderboard. It proves one thing — that OpenAI's own models got much better at everyday errands over the past year. It does not prove who is best at it, or how far anyone is from "no human supervision needed."
5. The same day, another test gave 12%
Interestingly, on the very day LifeVal was published, another organization handed out a much colder report card. SemiAnalysis released GDPx the same day: 206 long professional tasks across 11 high-value domains, each up to eight hours long, testing genuinely professional work. The result was a 12.0% pass rate. Same day, same kind of machine — 80.6% on everyday chores on one side, 12% on long professional work on the other. Put the two together and you get a complete picture of what agents can actually do today.
That contrast is exactly where LifeVal is most valuable. It draws a line: a thirty-minute errand and an eight-hour project are two entirely different kinds of difficulty. Finding a home, ordering groceries, unsubscribing from mailing lists are tasks with clear goals, limited steps, and recoverable mistakes. An eight-hour professional job needs sustained judgment, memory across steps, and self-correction after errors — if any link breaks, the whole thing can fall apart. Today's agents have got their hand on the doorknob of the first kind; the second is still far away.
LifeVal did not come out of nowhere, either — it has a predecessor. OpenAI already had GDPVal, a benchmark for professional assignments. LifeVal and GDPVal together form a complete "work plus life" yardstick: the eight hours at the office belong to GDPVal, the chores after work belong to LifeVal. Publishing both on the same day says something about OpenAI's ambition: it no longer wants only to "help you write code," it wants to "help you live your life." There is nothing wrong with that ambition, but it raises the stakes for evaluation — a mistake on a work task can be redone, while a mistake on a life task sometimes cannot be undone.
For vibe coding, this conclusion transfers directly. Building a landing page, fixing an error, adding a small feature — those are thirty-minute errands, and models can deliver them. Building a multi-module product from scratch and keeping the architecture consistent over days is an eight-hour project, and it still needs a human watching. LifeVal's 80.6% and GDPx's 12% together describe the real waterline of AI programming today: the easy jobs keep getting easier, and the hard jobs still need a human to cover them.
6. What the test really changes is what gets tested
The most important thing about LifeVal is not the score. It is that it pushes the definition of a "good model" one step forward. For the past few years, model quality has been defined by exam results: math, code, reasoning, one multiple-choice question after another. Now OpenAI says what really matters is whether a model can help ordinary people get through their days. That shift matters more than any single score, because it decides where labs will put their effort next year: browser use, computer use, and connectors — the ability to actually do things — are moving from add-on features to the main event.
But the shift also brings new risks. Everyday chores are not exam questions, and the cost of getting them wrong is real: pick the wrong kindergarten and a child loses a semester; fill in a tax form wrong and a fine arrives in the mail; get a prescription wrong and the consequences are worse. Scores can take partial credit; life cannot — either the job got done or it did not. LifeVal's gentle "average reward" arithmetic hides exactly that point: 90% on a task sounds good, but if the missing 10% is "don't get the medication name wrong," then the 90% is worth nothing. If future everyday-life benchmarks want to be trusted, they need to learn to score the way life itself scores: get a critical step wrong, and it is zero.
Three things are worth watching next, and all three are concrete. First, whether OpenAI publishes the task set and the grading rubric so others can rerun the test and compare — without that, it stays an internal report forever. Second, whether an independent party runs the same tasks against other labs' models, filling in the missing half of the leaderboard. Third, whether LifeVal moves from "showing progress" to "driving the product": if ChatGPT is really going to take on everyday chores by this standard, then error handling, human confirmation on critical steps, and accountability when things go wrong all have to catch up. Scores can rise thirty points in a year; trust is built much more slowly.
To sum up: LifeVal proves that AI is getting fast at everyday errands, and it also proves that the yardstick for measuring that ability is itself immature. 80.6% is a number worth remembering — but what is more worth remembering is the reminder behind it. As AI moves from the exam hall into everyday life, what matters is no longer how many questions it answered correctly, but which single time it messed up.
Comments (0)
Sources
Related articles
On 18 November 2026, Google Research and co-authors posted BuggyMLE to arXiv: bugs injected into 222 working ML repositories, and 14 models slide from 95% on the easiest tier to 9% on the hardest. A 9-billion-parameter model trained easy-to-hard beats Claude Opus 5.5 on the 10-bug tier.
On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.
On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.