OpenAI Unveils LifeVal, Its Benchmark for AI Agents on Everyday Life Tasks
On November 2, OpenAI unveiled LifeVal, an internal evaluation of daily tasks (childcare, administrative paperwork, budgeting, grocery shopping) that it has used for a year. It only published select scores for GPT-6 Astra, without an overall score or comparisons.
On November 2, 2026, OpenAI introduced LifeVal, an evaluation measuring the ability of its models to handle everyday life tasks: finding daycare spots, accessing public services, budgeting, ordering groceries, or planning a fitness routine. The company says it has used it for a year to monitor and improve its models, but it only released selected scores on a task-by-task basis.
Tasks Drawn from Time-Use Surveys
To select tasks, OpenAI explains that it started from two major surveys on how people spend their days: the American Time Use Survey and the Multinational Time Use Study. It supplemented these with a new “demographically representative” survey of 3,995 people, asking about the tasks for which they would like help from a personal assistant. The selected areas include: caregiving, health, education, household administrative tasks, personal finance, shopping, and accessing public services.
The goal is not to judge a chatbot’s answer, but an action seen through to completion. OpenAI writes that a truly helpful assistant must be able to take action so things get done, rather than merely offering advice, linking LifeVal to training its models on web browsing, computer use, and connectors.
Showcased Scores
The page highlights several tasks along with scores for OpenAI’s current models:
| Task (OpenAI examples) | Displayed score |
|---|---|
| Finding housing based on budget, independent unit, and commute | 100% |
| Organizing volunteers ahead of a flood | 98.0% |
| Setting up a game to watch remotely with a grandmother | 96.3% |
| Enrolling a child in the right preschool after a move | 93.1% |
| Meal planning for a recently diagnosed diabetic parent | 90.4% |
Scores published by OpenAI on its LifeVal page, accessed on November 2, 2026.
A second table lists tasks (booking a sensory workshop for a toddler, scheduling car maintenance, ordering halal groceries within a budget, unsubscribing from mailing lists…) where the best recorded score for GPT-6 Astra is 100%.
What Is Missing
The page accessed provides no overall score, no total number of tasks, no results for competing models, no detailed grading methodology, and does not indicate whether the tasks will be released publicly. The table refers to the “best recorded score” without specifying how many attempts were made per task. Another nuance: on X, OpenAI mentions feedback from “thousands of ChatGPT users,” whereas the page describes a representative survey of 3,995 people. LifeVal arrived the same day as SemiAnalysis’s GDPx agent benchmark, where no model exceeded 12% on expert tasks.
Our take
LifeVal asks a solid question: what can an AI assistant really do for everyday users, beyond standardized tests and coding competitions? Grounding the benchmark in time-use surveys is a serious approach. However, in its current form, it serves primarily as a showcase: OpenAI grades its own models on its own benchmark and exclusively displays successes, often at 100%. Without an aggregate score, competing models, or public tasks, it is impossible to know whether ChatGPT handles daily life better than Gemini or Claude, or how it performs on tasks that were not highlighted. The contrast with GDPx, published on the same day by a third party, underscores the gap between evaluations designed by a model developer and those measured externally.
Sources
- OpenAI — LifeVal (November 2, 2026) Primary source
- OpenAI on X — LifeVal announcement (November 2, 2026) Primary source
- SemiAnalysis — GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains (November 2, 2026) Primary source — For the comparison with GDPx
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
