OpenAI LifeVal: everyday-life agent eval spanning childcare, health, finances and government services

Why it matters

【Key points】

• OpenAI published LifeVal (2026-11-02): an evaluation of model performance on everyday life tasks (childcare, health, finances, government services, etc.), grounded in ATUS/MTUS time-use data plus a demographically representative survey of 3,995 respondents.
• Tasks emphasize completing workflows via browser/computer use and connectors, not just chat advice.
• Example scores on current models are reported in the high 90s for several vignettes.
• Positions LifeVal as a complement to academic benchmarks for measuring AGI usefulness in daily life.

【Our read】

Background

• frontier benches still skew MMLU/coding/math and tool quizzes, while consumer-assistant narrative is stuck on finishing real life errands.
• LifeVal anchors tasks in time-use surveys and scores closed-loop browser/computer/connector workflows—an explicit product metric for ChatGPT’s agentic path.

• Compare: GAIA/BrowseComp stress retrieval and multi-step correctness; AndroidWorld/OSWorld stress OS GUI; HELM is multi-axis profiling.
• LifeVal’s difference is “daily value” topic choice and “get it done” closure, not another knowledge leaderboard.
• Public material is still blog-level: incomplete reproducibility vs GAIA-style open protocols
• treat 90%+ vignette scores as OpenAI’s internal KPI made public, not a settled industry standard.

Impact

• short-term lifts “everyday agent eval” discourse and gives OpenAI a citeable KPI story for Computer Use/connectors
• watch for open protocols, third-party replication, and peer responses from Anthropic/Google.
• Ship as paper/news, not a model release.

【Verifiable】

• Primary(OpenAI, 2026-11-02). Claims ATUS/MTUS grounding and n=3,995 survey; vignette scores attributed to “latest models” in the post. Full public dataset/protocol/baselines across vendors not clearly released in the blog post