Does OpenAI’s LifeVal show AI can manage everyday life?
No. OpenAI reports that its latest internal model scores 80.6% on an evaluation inspired by everyday tasks, up from 49.3% for GPT-5.2. But it has not published the task denominator, full scoring protocol, uncertainty or evidence that the systems work reliably in people’s homes and services.
The short version
- What happened
- OpenAI reports an average unweighted LifeVal reward of 80.6% for GPT-6 Astra, compared with 49.3% for GPT-5.2, but the company page does not publish the number or complete list of evaluated tasks.
- Who it affects
- LifeVal recognises that unpaid care, household administration and access to services deserve attention, but the published score does not show that people receive safe or reliable help.
- What remains unproven
- OpenAI designed, used and reported the evaluation; no independent replication or external audit is provided.
Keep in mind
OpenAI designed, used and reported the evaluation; no independent replication or external audit is provided.
The public page omits the total task denominator, complete task set, full scoring protocol, judge details, trial counts, variance and confidence intervals.
The 3,995-person survey informed priorities rather than model performance, and its geography, recruitment, field dates, response rate, weighting and questionnaire are not disclosed.
An average unweighted reward can hide task-level failures and does not show whether small differences between models are statistically or practically meaningful.
Constructed task performance does not measure execution, long-term outcomes, privacy, recovery from errors or the distribution of benefits and harms in real settings.
Source analysis · 3 linked sources. Primary analysed source date: 2 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 10 min The evidence, methods, implications and limitations
The answer is no: LifeVal measures an internal test, not everyday reliability
OpenAI’s new LifeVal page makes a striking claim about progress: the average unweighted reward rises from 49.3% for GPT-5.2 to 80.6% for GPT-6 Astra on tasks intended to resemble things people need to do outside work. Intermediate models are listed at 73.0% for GPT-5.4, 77.4% for GPT-5.5 and 80.0% for GPT-5.6. The direction is clear within the company’s own evaluation. What is not clear is how much confidence readers should place in the level of those scores or how well they predict performance in a real household, classroom, clinic, financial decision or government service.
The page does not disclose the benchmark’s total number of tasks, publish the complete task set, describe every scoring rubric or identify how many trials produced each result. It provides no confidence intervals, variance, subgroup results or independent replication. OpenAI also says it has used LifeVal over the past year to track and drive model improvement. That makes the evaluation useful as an internal development instrument, but it also raises a familiar benchmark question: repeated exposure can improve performance on the instrument without proving the same gain on unseen situations. The responsible conclusion is therefore narrow. OpenAI reports substantial improvement on its own test; LifeVal does not by itself show that an AI agent can safely manage people’s lives.[1]
What LifeVal is trying to measure
Most prominent AI evaluations concentrate on coding, mathematics, science or office work. LifeVal starts from a different question: which activities consume people’s time and which of those could plausibly benefit from assistance? OpenAI says it drew on the American Time Use Survey and the Multinational Time Use Study to build an initial map of activities. The US Bureau of Labor Statistics describes ATUS as measuring time spent on work, childcare, volunteering, socialising and other daily activities. MTUS brings together more than one million diary days from more than 100 randomly sampled national-scale surveys, harmonised across 30 countries and 55 years.
Those datasets give the project a stronger foundation than a brainstormed list of chores, but they do not validate the benchmark. Time-use diaries describe what people do and for how long; they do not say which parts should be delegated, what a correct AI response looks like or which errors are acceptable. OpenAI groups LifeVal tasks across caregiving, health, education, household administration, personal finance, shopping and government services. These categories mix low-stakes planning with decisions that can affect health, money, legal position and access to support. A single average score can therefore hide very different consequences when a system fails.[1][2][3]
The distinction that matters
What LifeVal can and cannot establish
The company combines public time-use evidence, a priority survey and internal task scoring, but only the last step produces the model percentages.
Activity map
Use ATUS and MTUS to identify everyday activities
Shows what people do, not whether AI should do it
Priority survey
Ask 3,995 people where help could matter
Ranks interests; does not test model use
Internal evaluation
Score selected models on constructed tasks
Reports 49.3% to 80.6% average reward
Not established
Deploy with people in real settings
No independent outcome, safety or equity trial reported
The 3,995-person survey is a priority-setting step, not a performance trial
OpenAI says it ran a demographically representative survey of 3,995 people to learn which everyday activities matter most and where respondents would value help. That is a meaningful denominator and a welcome attempt to ask people rather than assume what they need. It should not be confused with 3,995 people using an AI agent. The survey helped prioritise domains and tasks; it did not test whether the model completed them successfully for those respondents or improved their outcomes.
Important survey details are absent from the public page. It does not state which country or countries the respondents represented, how they were recruited, when the fieldwork happened, what response rate was achieved, whether results were weighted, how demographic representation was defined or how questions were worded. Those details affect interpretation. People with disabilities, carers, older adults, people navigating unfamiliar public systems and households with limited digital access may place different value on assistance and face different costs from mistakes. Without the questionnaire and subgroup results, the survey supports the claim that OpenAI sought public priorities, but not that the resulting benchmark represents every population equally.[1]
What the published examples show—and what they leave out
The LifeVal page presents concrete examples. It reports scores of 93.1% for planning activities for a kindergartener, 98.0% for coordinating volunteers after a flood, 100% for finding a home that meets stated requirements, 96.3% for arranging to watch a game remotely and 90.4% for creating a diabetes meal and care plan. Examples make the intended scope legible. They do not reveal whether each item was scored once or repeatedly, whether a person, model or deterministic rule judged the answer, which factual tools were available, or what counted as partial credit.
A high percentage also has different meaning across tasks. A home search could be checked against a fixed list of constraints, although real listings change and affordability judgments are personal. A diabetes plan can be medically consequential and depends on diagnosis, medicines, culture, budget and clinical advice. Flood coordination involves changing conditions and responsibility for volunteers. If the benchmark grades the quality of a written plan, it may test reasoning and instruction-following without testing execution, consent, privacy, tool failure or consequences. OpenAI calls the metric an average unweighted reward, but the public description does not explain whether every task contributes equally or how safety-critical mistakes affect the result.[1]
Why a rising internal score can still be useful
The missing details do not make LifeVal worthless. A stable internal evaluation can help a laboratory identify regressions, compare models and direct engineering attention toward tasks that ordinary users recognise. The progression reported across five model generations is large enough to justify further examination. The task examples also challenge an evaluation culture that treats professional productivity as the only valuable outcome. Care work, household administration and access to services are economically and socially important even when they are unpaid.
But a development benchmark has a different evidentiary role from an external audit. OpenAI designed the evaluation, selected the models, used it during development and published the comparison. The page does not show performance from rival systems under the same conditions, blinded human baselines or independent reproduction. Nor does it report the distribution of errors. An 80.6% average could represent broadly competent performance, very strong results on easy tasks with serious gaps elsewhere, or variability that disappears in the mean. Without task-level denominators and repeated trials, readers cannot calculate how often a particular kind of request fails or whether the reported 0.6-point difference between GPT-5.6 and GPT-6 Astra is meaningful.[1]
What this means for families, professionals and public services
For a family, the result is not a reason to hand over medication, benefits, budgeting or care decisions without checking. A model may produce a polished plan while overlooking an allergy, eligibility rule, local deadline or personal constraint. Teachers and clinicians should read the examples as research prompts, not evidence that a general-purpose assistant is validated for students or patients. Public-service teams should require task-specific testing, accessibility review, privacy protections and a route to a responsible human before an agent can act on someone’s behalf.
The benchmark also points to a real opportunity. Good assistance with forms, comparisons, scheduling and coordination could return time to people already carrying unpaid administrative and care work. Benefits will depend on more than model capability: services must provide accurate structured information, people must be able to correct the agent, and responsibility for errors must remain clear. If an agent fails to submit a claim or gives unsafe health advice, the harm belongs to the person even when the benchmark records only a lower reward. Real evaluation must therefore measure completion, recovery from mistakes, human effort, distributional effects and the outcomes people actually value.[1]
What evidence would change the assessment
OpenAI could make LifeVal substantially more informative by publishing the task count and taxonomy, versioned prompts, scoring rubrics, judge identity, tool environment, model settings, trial counts and task-level results. It should report uncertainty, failure categories and performance across survey-defined groups, while protecting participants. A held-out set that developers cannot inspect would help separate genuine generalisation from optimisation to the benchmark. Comparisons with humans and other systems under identical conditions would establish a more useful reference point.
The decisive next step is prospective, independent testing with consenting users. Studies should prespecify tasks and outcomes, compare AI assistance with existing support, record the time people save or spend correcting mistakes, and track safety incidents, privacy failures and unequal performance. High-stakes domains need professional review and longer follow-up. Until that evidence exists, LifeVal is best read as a transparent signpost toward a broader idea of useful AI—and an incomplete company benchmark whose top-line score cannot carry the weight of real-world proof.[1][2][3]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 2 min 31 sec · Background explainer
What is artificial intelligence? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A short introduction to artificial intelligence from the Royal Society. Start here for basic concepts, then use current reporting to explore what has changed.
Evidence check: Published in 2018, before today’s generative AI tools. Useful for basic concepts, rather than an overview of current capabilities.
Published 2018-11-28. Source checked 14 October 2026. Check captions on the publisher’s video page.
Explore current AI technologyPhoto: ThisisEngineering / Unsplash · Licence
Continue the story
Read next
Continue exploring ai agents: compare the finding and its evidence limits.
An AI agreed with one senior investor 88.6% of the time—what does that prove?
Terminal X and OpenAI report that an evaluation agent matched one senior reviewer across 500 scored judgments and caught more mandate drift than the next-best unnamed model. It is useful company evidence, not an independent test of investment performance.
Read the next reportExplore the full source and evidence record
What this means for people
- LifeVal recognises that unpaid care, household administration and access to services deserve attention, but the published score does not show that people receive safe or reliable help.
- Errors in health, finance or government-service tasks can cost money, time, care or legal rights. Users and professionals still need task-specific safeguards and accountable human review.
Global context
LifeVal uses a US time-use survey and a harmonised international resource spanning 30 countries, but the public page does not identify the geography of its 3,995-person priority survey or report country-level model results. Activities, institutions and acceptable assistance vary across languages, cultures, health systems and public services. International source material therefore broadens the activity map without proving that the benchmark or its scores transfer globally.
What the evidence does not yet show
- OpenAI designed, used and reported the evaluation; no independent replication or external audit is provided.
- The public page omits the total task denominator, complete task set, full scoring protocol, judge details, trial counts, variance and confidence intervals.
- The 3,995-person survey informed priorities rather than model performance, and its geography, recruitment, field dates, response rate, weighting and questionnaire are not disclosed.
- An average unweighted reward can hide task-level failures and does not show whether small differences between models are statistically or practically meaningful.
- Constructed task performance does not measure execution, long-term outcomes, privacy, recovery from errors or the distribution of benefits and harms in real settings.
What to watch next
- Publication of the complete task denominator, taxonomy, prompts, scoring rules and version history
- Independent reproduction with task-level uncertainty and comparable human and model baselines
- Survey methods and subgroup findings for the 3,995-person priority study
- Prospective trials measuring successful completion, correction effort, safety, privacy and unequal performance
- Evidence that gains persist on held-out tasks not used to guide model development
Living evidence record
Impact record IAI-0M4JZHK
Evidence stage
Announced
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
3 November 2026
Source trail
3 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 3 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 3 November 2026
Research question, topics and regions
Whether an internal benchmark establishes that AI systems can reliably complete consequential everyday tasks
Topics: AI agents · Benchmarking · Everyday tasks · Model evaluation · Time-use research
Regions: United States · International
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.