Stubbi
Rankings
Benchmark

LifeVal

Test whether an AI assistant can complete everyday tasks using a person’s preferences and context.

Checked Nov 17, 2026OpenAI paper and methodology

Explore the original evidence

Explore the original evaluation, its methods, and reported findings.

Open eval

Coverage

169 tasks selected using a survey of consumer priorities, spanning areas such as household administration, caregiving and shopping. OpenAI. Paper authors: Jenny Yang, Elizabeth Proehl, Anjan Bharadwaj, Emily Fan, Charlotte Cole, James Kim, Noam Brown, Phoebe Thacker, Amelia Glaese and Tejal Patwardhan.

Methodology

Agents work in task environments with an offline web browser. A rubric grader checks artifacts and changes to the environment; a separate grader rejects runs that exploit internal environment APIs. The reported main grader uses GPT-5.6 Sol at xhigh reasoning. The survey is US-centered and willingness to pay is a stated preference. The browsing cache can change between runs. These are publisher-reported evaluations, not independently reproduced results.

People & organizations

1

Sources

HN discovery discussion