What happened
OpenAI published LifeVal on November 2, 2026, describing it as an evaluation for measuring progress on tasks that matter in everyday life. Its examples span caregiving, health, education, household administration, personal finances, shopping, and accessing government services. OpenAI says it built the evaluation to track and drive model improvement over the past year, and argues that a personal assistant needs to do more than provide advice or a chat response: it should be able to take actions that get things done.
The task set began with two sources on how people spend their time: the American Time Use Survey and the Multinational Time Use Study. OpenAI says these datasets helped identify everyday tasks across demographic groups, times of day, and seasons. It then conducted a new demographically representative survey of 3,995 respondents, asking where people most wanted help from AI and personal agents. The company says the resulting tasks reflect the areas where people wanted assistance most; the source does not provide a detailed breakdown of survey responses by task or demographic group.
OpenAI presents an average-reward chart for five model releases, with the model release dates shown as December 11, 2025; March 5, 2026; April 23, 2026; July 9, 2026; and September 3, 2026. The recorded scores rise from 49.3% for GPT-5.2 to 73.0% for GPT-5.4, 77.4% for GPT-5.5, 80.0% for GPT-5.6, and 80.6% for GPT-6 Astra. OpenAI labels the measure “Average reward (unweighted).” The article says that over the past year its models were trained to become faster, more capable, and more aligned on browser use, computer use, and connector tasks, and that enabling these abilities expands the work ChatGPT can help with.
The page also gives individual task examples rather than only an aggregate score. It reports 93.1% for finding and shortlisting kindergartens after a move, 98.0% for organizing flood-preparation volunteers, 100% for finding a family home matching several constraints and adding a deed to a records cart, 96.3% for arranging a remote game watch with a grandparent, and 90.4% for planning meals and care after a parent’s diabetes diagnosis. A table lists several additional tasks at 100%, including booking a sensory class, arranging a car inspection, ordering groceries within a budget, and unsubscribing from mailing lists. The source gives no price or separate access mechanism for LifeVal; it presents the evaluation as a way to measure model progress.
What the archive adds
An earlier Telegram post covered OpenAI’s GDPval, a benchmark organized around occupational tasks. LifeVal is a new event, not a repeat of that announcement: its stated focus is everyday personal-assistant workflows selected with input from a representative survey, rather than professional roles. The supplied archive has no related Field Note URL to link here.
Trend signal
No supplied SITE_PATTERNS thesis directly fits this announcement. LifeVal evaluates practical tasks and reports model scores, but the source does not establish a shift in how agent systems are integrated, governed, or evaluated across the industry.
Why now
OpenAI says it used LifeVal to track and drive improvement over the past year, and frames the results alongside progress in browser use, computer use, and connectors. The source does not explain why the benchmark is being published on this particular date.
Why it matters
For teams building assistants, LifeVal offers a useful reminder that user value often depends on a chain of actions—researching options, respecting constraints, preparing an artifact, and leaving consequential choices for review—not on a fluent answer alone. OpenAI’s examples make those workflows legible, and its reported improvement suggests that the company sees browser, computer, and connector capabilities as part of the path toward more useful assistance.
The scores are not, by themselves, a deployment decision. A team should translate its own high-value workflows into explicit tasks, define what counts as a successful outcome, and test whether the system handles exceptions and asks for approval at the right points. LifeVal can help broaden the questions teams ask, but the supplied results do not show whether a model will be reliable on a particular company’s tools, policies, or users.
Limits and uncertainty
The announcement does not provide the full task set, the evaluation rubric, reward calculation details, or results for every model and task. It does not explain how the aggregate score maps to real-world success, whether tasks were run with identical tools and conditions, or how human review and error costs were represented. The survey is described as demographically representative, but its sampling and question wording are not detailed in the supplied material. The examples are selected by OpenAI, and the results are company-reported; they do not establish independent replication or comparative performance against other providers.
Reader takeaway
LifeVal makes everyday assistance—not just knowledge recall—a measured target for model progress. Its scores and workflow examples are worth treating as a prompt for better evaluations, not as proof that an assistant can reliably handle a team’s real tasks. Builders should define the outcome, constraints, and review points that matter in their own workflows and test them directly.