All Radar picks

X OPPORTUNITY RADAR / THE BRIEF

OpenAI Gave Its Agent 100% for Shopping for a House

LifeVal measures AI on childcare, groceries, government forms and caregiving. The examples are vivid. The public benchmark is not independently runnable.

Information Overload
2 min readSources
AI-generated conceptual editorial illustration of a polished robot holding a giant gold check mark in front of a mirror while a family watches from beside groceries, childcare blocks, medicine, paperwork, house keys and a teacup.
AI-generated conceptual editorial illustration. It satirizes an institution grading its own agent on everyday tasks; it is not a LifeVal interface, a benchmark result screen or documentary evidence of task performance.

OpenAI’s new LifeVal evaluation asks agents to handle the administrative obstacle course that fills ordinary life: finding childcare, navigating government services, managing a budget, ordering groceries, supporting a relative and preparing for school.[1]

That is a useful direction for AI evaluation. People do not experience a model as a leaderboard entry. They experience it when the form is wrong, the appointment is missed or the supposed shortcut creates another hour of cleanup.

A 100% score with a smaller meaning

OpenAI’s public examples include 93.1% for shortlisting kindergartens, 98% for organizing flood-preparation volunteers, 96.3% for arranging to watch a game remotely with a grandmother, 90.4% for a diabetes meal plan and 100% for finding a new home.[1]

The house example is narrower than the headline number sounds. The task asked the agent to find a home matching constraints, calculate transfer tax and add a deed to a county-records cart for review. A perfect benchmark score did not mean the model bought a house, cleared title, negotiated a mortgage or discovered every expensive surprise behind the drywall.

OpenAI says it began with the American Time Use Survey and the Multinational Time Use Study, then conducted a “demographically representative” survey of 3,995 respondents to identify areas where people wanted help. The resulting categories include caregiving, health, education, household administration, finance, shopping and public services.[1]

The public page is a showcase, not a reproducible evaluation

OpenAI does disclose more than the examples: its chart covers 169 tasks, names the models compared and explains how it averages attempts and tasks. GPT-6 Astra at xhigh records 80.6% average reward, while the expanded examples list best-recorded task scores. But the announcement does not link a downloadable task set, evaluation code or grader. It also provides no competitor results or independent replication. Those disclosures help readers interpret the results without making the evaluation independently reproducible.[1]

That does not make the scores false. It limits what outsiders can conclude from them. A vivid example is evidence that OpenAI chose and demonstrated that workflow under its own evaluation. It is not yet a public instrument that another lab, consumer group or regulator can rerun on equal terms.

LifeVal’s cultural significance is the premise: agentic AI is moving from answering questions into touching the errands, records and relationships that hold a household together. That raises the value of the benchmark and the cost of grading it behind glass.

If everyday life is going to become the test environment, everyday people deserve more than a highlight reel. Publish the tasks, the grader, the failures and the conditions. A household is not a demo account, even when the score says 100%.

CORRECTIONS & UPDATES.

Correction ·
Corrected the claim that task counts and model details were undisclosed. The source includes 169 tasks, aggregation methodology and model results; displayed examples are best-recorded task scores.

FOLLOW THE SOURCES.

Checked Nov 2, 2026. Numbered references point to the evidence behind factual passages. Commentary is our own.

  1. LifeVal — November 2, 2026 (opens in a new tab)OpenAIBack to cited passage
Artificial intelligenceBenchmarksAI agentsEveryday lifeEvaluationOpenAI
Back to the latest picks