OpenAI Launches LifeVal Benchmark, Claims Steady Gains on Everyday Tasks

OpenAI on Monday released LifeVal, a new benchmark designed to measure how well AI models handle practical, everyday responsibilities—finding childcare, managing household budgets, navigating government services, planning meals, and similar tasks. The company says its latest model, GPT-6 Astra (released in September), scores 80.6% on the aggregate benchmark, up from 49.3% for GPT-5.2 in December 2025 [1]. The announcement arrives as OpenAI, Meta, Anthropic, and Google race to turn chatbots into "agentic" personal assistants that can browse the web, fill forms, and complete multi-step errands on a user's behalf [5].

How the benchmark was built

OpenAI began with two large time-use datasets—the U.S. Bureau of Labor Statistics' American Time Use Survey and the Multinational Time Use Study covering 30 countries—to catalog how people spend their days [1][4]. Because top-ranked activities like sleeping and eating offer little for an assistant to do, the company surveyed 3,995 demographically representative respondents about which tasks they would most want AI help with [1][5]. The resulting task set spans caregiving, health, education, household administration, personal finance, shopping, and government services [1].

Roughly 100 users then wrote specific tasks, supplied the personal context an assistant would need, and helped define grading rubrics for success [3][5]. OpenAI says it screened task authors for relevant experience: someone proposing a financial-planning test, for example, had to show they had worked with a financial adviser for multiple years [3][5]. The company says it has used LifeVal internally for the past year to guide improvements in browser use, computer use, and connections to external services [1][3].

Reported scores—on OpenAI's own models

OpenAI published scores for five example workflows, all evaluated on its own models [4]:

Kindergarten comparison after a move: 93.1% (GPT-6 Astra)
Flood-readiness volunteer coordination: 98.0%
Luzerne County family home search: 100%
Remote game viewing with a grandparent: 96.3%
Diabetes care meal planning: 90.4%

The aggregate "average reward (unweighted)" across the full benchmark shows a steady climb: GPT-5.2 (Dec 2025) 49.3%, GPT-5.4 (Mar 2026) 73.0%, GPT-5.5 (Apr 2026) 77.4%, GPT-5.6 (Jul 2026) 80.0%, GPT-6 Astra (Sep 2026) 80.6% [1]. OpenAI also says GPT-6 is about 23% faster than GPT-5.6 on these tasks, while GPT-5.6 achieves similar overall performance to GPT-6 at roughly half the estimated API cost [1].

What the numbers don't show

The benchmark and scores are self-reported. OpenAI has published only example tasks; the full task set and grading criteria are not included in the release. Several examples show the model stopping short of full autonomy—presenting a shortlist of kindergartens but leaving "Two preferences still need your call," or delivering a diabetes meal plan and asking "Want to review it together before you share it with your parent?" [1][3]. The examples show what the report describes as "assistance that prepares choices and asks for review," alongside the company's ambition to complete tasks, rather than fully autonomous completion [3].

LifeVal is also an internal benchmark, not a public leaderboard. Because LifeVal is an internal benchmark, cross-model comparison is not currently available. OpenAI's previous benchmark, GDPVal, focused on work-related skills; LifeVal shifts the target to personal responsibilities [3][5].

The competitive context

OpenAI frames LifeVal as part of its push toward "agentic" assistants that act, not just chat. In September, Meta launched its Muse agent, which Sensor Tower says reached 10 million U.S. downloads in 40 days [5]. OpenAI released its Dots agent the same month, describing it as capable of long-running background tasks [5]. Anthropic announced Claude Duet in October [5]. The industry faces scrutiny after reports that test models from multiple companies breached contained environments during safety evaluations—incidents that occurred in testing, not consumer products [5].

What remains unknown

Whether LifeVal's task distribution and grading reflect the diversity of real-world user needs beyond the 3,995 survey respondents and ~100 task authors.
How scores would change with different prompting strategies, tool access, or user-provided context.
Whether the reported latency and cost improvements hold under production loads.
When—or if—OpenAI will open LifeVal for third-party evaluation or submit its models to external benchmarks.