OpenAI launches LifeVal, a test of how well AI handles everyday chores

OpenAI on Monday, November 2, 2026, introduced LifeVal, a benchmark that measures how well AI models handle everyday tasks such as finding childcare, creating a budget, ordering groceries and accessing government services. In OpenAI’s own testing, its GPT-6 Astra model scores 80% on general performance across these everyday-task prompts, while GPT-5.2 from December 2025 scores 49.3%.

The benchmark shifts the yardstick away from exam-style knowledge tests and towards whether a model can actually get errands done. OpenAI frames it as a way to measure “what matters in the lives of everyday people” rather than “academic assessments of model knowledge”. All of the published numbers are OpenAI scoring its own models, and no outside group has yet run the test.

What LifeVal measures

OpenAI says that for the more than one billion people it reports use ChatGPT every week, it wants AI that helps with “the practical challenges of everyday life”. Its examples include finding childcare, accessing government services, creating a budget, preparing for school, modifying recipes, ordering groceries and designing workout plans. The task set spans caregiving, health, education, household administration, personal finances, shopping and accessing government services.

The emphasis is on action, not advice. OpenAI writes that to be “a truly helpful personal assistant, ChatGPT should be able to take actions that actually get things done”. According to CNN, LifeVal evaluates the skills a model needs to reach the right answer: how well it navigates websites, how well it applies context by looking through large amounts of information in a user’s chat history, and whether it reaches conclusions that match the user’s needs and expectations.

OpenAI says it has used the evaluation over the past year to “track and drive improvement in model performance on everyday tasks”. In other words, LifeVal is not only a scoreboard released alongside a model; it has also shaped how OpenAI trained those models. The company says that over the same period it trained its models to become “much faster, more capable, and more aligned” on browser use, computer use and connectors tasks (connectors are integrations that let ChatGPT reach outside apps and data).

How OpenAI built the task list

OpenAI started from two large studies of how people spend their time: the American Time Use Survey and the Multinational Time Use Study. CNN notes a snag with that starting point: some of the most time-consuming activities, such as sleeping and eating, are not things AI can help with.

To find where people actually want AI help, OpenAI ran what it describes as a new, demographically representative survey of 3,995 respondents. CNN puts the figure at “around 4,000” and adds a second step: a smaller group of about 100 users wrote down tasks they wanted AI to do, the personal context they would give the model, and a rubric for judging whether the AI had succeeded.

OpenAI also screened those contributors for expertise. In CNN’s example, someone proposing a financial-planning task would need experience working with a financial advisor for multiple years. Jenny Yang, a researcher at OpenAI, described the bar to CNN:

“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.”

That quote appears in CNN’s report. Yang also explained the motivation: “Because so many people are using the app, we want to capture all the different use cases and the ways that they’re using it, and really understand like how we’re improving over time”.

How GPT-6 Astra scores against GPT-5.2

The headline comparison, as CNN reports it from OpenAI’s testing, is a jump from 49.3% for GPT-5.2 (December 2025) to 80% for GPT-6 Astra, which OpenAI released in September. Yang attributed part of the gain to handling large amounts of context: “There’s just so much information, and a better model would be able to sift through that quickly and really pull out the most relevant pieces”.

OpenAI’s announcement highlights individual tasks. The scores below are OpenAI’s own figures for its current models.

Task (OpenAI’s wording)	Score
Enroll in the right kindergarten after moving	93.1%
Manage volunteers ahead of a flood	98.0%
Find a new home for my family	100%
Watch a game with grandma from far away	96.3%
Meal prep for a parent adjusting to a new diabetes diagnosis	90.4%

CNN attributes the kindergarten (93.1%), flood volunteer (98%) and diabetes meal plan (90.4%) scores specifically to GPT-6 Astra. The home-search example asks the model to find a home in Luzerne County that fits a budget, has a separate living unit and keeps the commute short after school drop-off, then add the deed to a cart for review.

A second table on OpenAI’s page lists tasks under the heading “Best recorded GPT‑6 Astra score,” all at 100%:

• Find and book a sensory class that fits a toddler’s needs and the family’s schedule
• Check when a car is due for service and book an inspection before a road trip
• Order halal groceries for three favourite meals within a budget and itemize the order in a spreadsheet
• Find a replacement for a lost engagement ring using a photo and its diamond report
• Find and order one pocket-sized device to replace a phone, e-reader and handheld game console
• Unsubscribe from email lists

OpenAI summarises: “Our latest models score very highly on many LifeVal tasks”. The page does not list the full task set, the number of tasks, or how the 80% overall figure is aggregated.

How LifeVal fits OpenAI’s earlier benchmarks

LifeVal follows GDPVal, which OpenAI released last year to measure AI performance on work-related skills across industries such as manufacturing, government work, finance and sales. LifeVal moves the same idea from the office into the home.

OpenAI has a longer history of publishing evaluations. Its HumanEval dataset, introduced with the Codex coding model, consists of 164 hand-written programming problems checked by unit tests, an approach known as “functional correctness” that judges code by whether it works rather than how closely it matches a reference answer. OpenAI also released Evals, an open-source framework for creating and running benchmarks. LifeVal shares HumanEval’s idea of grading outcomes, but its rubrics come from users and many tasks have no single correct answer.

CNN points out why such tests matter beyond marketing: AI labs use standardized evaluations as guidance to train and optimize models in specific areas. A benchmark that rewards finishing errands will steer future models towards finishing errands.

What it means for builders

For developers building agents, LifeVal is a signal of where OpenAI is pointing its models. CNN describes the benchmark as part of OpenAI’s push to turn ChatGPT from a tool for generating content into an assistant that gets things done on users’ behalf. OpenAI says enabling browser use, computer use and connectors helps move ChatGPT “towards a truly agentic personal assistant”.

The skills LifeVal tests map directly onto what agent builders care about: reliable website navigation, pulling the relevant facts out of a long history, and matching a specific user’s constraints such as budget, schedule or dietary needs. The task examples also show the kind of personal context these agents are expected to handle, from a parent’s medical diagnosis to a family’s home search. Teams designing their own evaluations can borrow the method: ask real users what they want done, have experienced people write the task and the grading rubric, and score the outcome.

A September hands-on review of GPT-6 Astra is a reminder of the gap between scores and daily use. The tester notes that launch coverage focused on FrontierMath, GPQA and ARC-AGI results, which “tell you almost nothing” about whether the model will handle a messy spreadsheet reliably. In one test, Astra reported a job as done and delivered a zip file that would not open; an identical rerun produced six usable image files. The tester also found the model behind a “Work” toggle in ChatGPT, with a control that looked the same as the “Thinking effort” slider in Chat but selected the model instead, and noted that the Plus account defaulted to “Astra Light”. Those are single-user observations, not a benchmark, but they show the kind of failure a task score can hide.

Limits and open questions

Several points remain unclear from the sources:

• Self-reported results. Every LifeVal number comes from OpenAI testing its own models. No rival lab’s models appear in the published comparisons.
• Access. CNN calls LifeVal “a new internal tool”, and OpenAI’s page does not say the tasks or rubrics are released for others to run. Without that, outside researchers cannot reproduce the scores.
• Subjective grading. CNN notes that the steps for tasks like grocery shopping or choosing a kindergarten are “broad and vague,” results can be subjective, and an agent’s answers vary with the information it is given.
• “Best recorded” scores. The 100% table is labelled “Best recorded GPT‑6 Astra score”. The page does not say how many runs were made or how often the model succeeds on a typical attempt.
• Trained against the test. OpenAI says it used LifeVal to “drive improvement” over the past year. Part of the jump from 49.3% to 80% may reflect optimising for this specific evaluation, which makes independent tests more important.

The competitive backdrop is also intense. CNN reports that Meta, Anthropic and Google are all racing to build agents for real-world tasks, and that Meta’s Muse agent, launched in September, reached 10 million US downloads in its first 40 days, according to market intelligence firm Sensor Tower. A benchmark designed by one company, on which that company’s newest model leads, is a useful progress tracker for that company, not yet a neutral ranking.

For broader context on how third-party scores compare with makers’ claims, see Unicorn’s coverage of the refreshed AI hallucination benchmark and of GPT-Rosalind-6’s pricing relative to GPT-6 Astra.

What to watch next

• Whether OpenAI publishes the full LifeVal task list, rubrics and scoring method, or lets outside labs run it.
• Whether Meta, Anthropic or Google respond with everyday-task scores of their own, on LifeVal or a rival test.
• How OpenAI uses LifeVal results in future model launches, and whether the 80% overall figure keeps rising.
• Independent hands-on reports that check whether high task scores hold up for ordinary users with real errands.