OpenAI launches LifeVal, a benchmark that scores AI on everyday chores from childcare to government paperwork

What LifeVal measures and how it works

LifeVal evaluates AI agents on whether they complete everyday tasks people would want to delegate, rather than on whether they answer a question correctly. Tasks were selected from the American Time Use Survey and other sources, then validated through a demographically representative survey of 3,995 respondents who were asked which activities they would most want AI to handle, OpenAI said in its announcement.

The benchmark covers caregiving, household administration and government services. Example tasks include booking a toddler sensory class, finding a halal grocery order within a fixed budget and building a spreadsheet, and scheduling a pre-trip vehicle inspection. Each scenario is designed to require multiple steps — search, form completion, decision-making and verification — rather than a single retrieval or generation.CellCog reports that the resulting evaluation frames everyday administration as a structural AI challenge, not just a convenience.

"The idea is straightforward and useful: evaluate AI against the work people actually want help with," wrote Curtis Pyke in an analysis for Kingy AI. "The harder question is how much a displayed score tells us about completing that work consistently, with the right permissions and a result the person can rely on."

GPT-6 Astra’s published scores

OpenAI reported that GPT-6 Astra achieved a best-recorded score of 80.6% overall on the LifeVal benchmark. Three specific tasks scored 100%: toddler sensory-class booking, scheduling a pre-trip vehicle inspection, and budgeted halal grocery ordering with spreadsheet output.

However, OpenAI has not disclosed the number of runs, the scoring denominator, or how much scores vary across attempts. Kingy AI noted this limitation: "A percentage needs a denominator and a scoring rule before a reader can interpret it precisely. It might reflect completed trials, satisfied grading criteria, or another measure. A selected example is not an overall benchmark result."

The difference matters for practical use. As Kingy AI's analysis illustrates, even if each step in a ten-step workflow succeeds 95% of the time, the total chance of all ten completing correctly is roughly 60%. LifeVal's current reporting approach — best-score highlights without variance — makes it hard to assess how reliably a model would perform the same errand on a typical attempt.

OpenAI has not said whether LifeVal scores will directly influence GPT-7 development targets or product roadmaps. The company described LifeVal as an internal research tool developed over the past year that is not tied to any specific product launch, but the timing — concurrent with the broader industry push toward agentic AI products — suggests it is part of a deliberate effort to benchmark agent performance on precisely the kinds of tasks that could become near-term consumer use cases. Agents that can reliably complete government paperwork, manage household logistics and handle caregiving research would open new product categories beyond chatbots and coding assistants.

How LifeVal compares to other agent benchmarks

LifeVal sits alongside a growing set of agent-focused evaluations, though it differs in scope.WebArena introduced self-hostable websites for testing functional web tasks, and OSWorld 2.0 scored agents against fine-grained final-state checkpoints with partial credit. LifeVal prioritises real-world face validity — tasks people say they actually do — over controlled reproducibility.

A task in LifeVal can begin with a search, move to a form, pause for a human decision, and end days later with a confirmation receipt. That makes it harder to fully automate the evaluation but arguably more relevant to the consumer use cases OpenAI is likely to pursue. The company said LifeVal was developed internally over the past year and is not tied to any specific product launch.

What the benchmark does not measure

LifeVal scores task completion but does not currently report on several dimensions that matter for real-world delegation: how much time the AI saved versus doing the task manually, how often a human had to intervene, whether partial progress is counted, or what types of errors remain. OpenAI's system card for the earlier ChatGPT agent, which is a different system, discussed prompt-injection risks — a security dimension LifeVal does not directly evaluate.

The distinction between "completing a task" and "completing it safely, privately and without follow-up" is where the practical gap between a benchmark score and a useful assistant widens, as bothCellCog and Kingy AI note. A model that books the right appointment but exposes personal data in the process would satisfy LifeVal's current criteria.

Industry reception and open questions

Early reactions from the AI research community have been broadly positive about LifeVal's design philosophy but cautious about the lack of methodological transparency. “It is refreshing to see a benchmark built around what people actually spend their time on, rather than another multiple-choice dataset scraped from the web,” wrote one AI safety researcher on X who asked not to be named. “But any benchmark that reports a single best-score number without distribution or variance is more useful as a PR instrument than an engineering tool.” OpenAI has not released the full LifeVal task set or evaluation harness, which means other labs cannot independently reproduce the results or compare their own models on the same scenarios.