OpenAI introduced LifeVal on November 2, 2026, to evaluate AI help with everyday tasks. For households juggling shopping, school arrangements and paperwork, the results raise a useful question: what would count as a job actually done?
This is evaluation evidence, not a new consumer feature. For a household deciding whether to try AI assistance, the practical starting point is a small job with a result that can be checked.
What the scores do—and do not—show
OpenAI labels its table “Best recorded GPT-6 Astra score”. Six displayed tasks score 100%, including grocery ordering and car-service booking. A kindergarten example shows 93.1%, but illustrates a shortlist awaiting review, not completed enrolment.
Those are OpenAI’s reported results. The accessible announcement text does not define scoring or attempt counts. A best result cannot establish typical reliability; the percentages should not be read as a household’s probability of success. Judging repeatability would require repeated results and a clear definition of success.
OpenAI also cites a survey of 3,995 people to guide task selection. That count measures respondents, not completed AI tasks.
Check the product separately from the benchmark
OpenAI’s current cloud-browser documentation says ChatGPT Work can interact with supported public and signed-in websites. Access is available on paid ChatGPT plans, excluding Free and Go, subject to supported regions, rollout and workspace permissions. Some websites or transaction steps are unsupported; you may need to take over.
A managed work account has further limits: cloud-browser website sign-in is unavailable in Enterprise and Edu workspaces. Enterprise browser access also needs administrator enablement separately from access to Work.
Permission to visit a website is separate from approval for an important action. OpenAI says the browser is designed to seek confirmation before commitments such as a booking or payment. That describes the intended control, not a guarantee that every task will be handled correctly.
Usage is another constraint. ChatGPT Work and Codex share allowances, and consumption varies with the model, task, context and tools. Published message estimates are not a quota of successful household jobs. API token prices cannot tell you how many bookings a subscription will complete.
Define the finish line first
A useful way to apply these limits is to agree on the deliverable before delegating. A shortlist, a prepared form and a confirmed booking are different deliverables. For a first trial, choose one of these stopping points:
Research. Ask for options with source links and unresolved questions. For school choices, check whether the listed hours and locations match the sources; the household still decides which option fits.
Preparation. Ask for a draft or ready-to-review form, with missing information flagged. Treat this as work prepared for you, not something already submitted.
Completion. If an action is authorized and supported, check the provider’s confirmation or account record. For an appointment, look for the confirmed date and time—not just an assistant saying it has finished.
Keep that endpoint separate from permission to act. RohitAI’s explanation of OpenAI’s rules on personal context and permission explores those boundaries in more detail.
Judge the help by the time you get back
To assess a trial, count your attention, not simply how long the AI runs. A practical measure is the time the task would otherwise take you, minus the time spent giving instructions, checking and correcting the result. This is a proposed assessment, not a measured saving; it only compares like-for-like work at an acceptable quality.
A handoff for personal judgment need not make assistance useless. The question is whether the agreed work was done correctly with less effort from you. Start there before expanding the task or paying for more capacity.