OpenAI launches LifeVal: a test of AI agents in everyday life

What's new

On November 2, 2026, OpenAI published a new test called LifeVal. It measures whether AI agents can carry out everyday tasks. An AI agent is a program that doesn't just answer questions but gets started on its own and gets things done.

The tasks require action rather than advice. They might include finding child care, booking a vehicle inspection and appraisal, ordering groceries or handling government paperwork. The tasks are divided into seven categories: caregiving, health, education, household administration, personal finance, shopping and public services.

According to an independent blog, OpenAI has used LifeVal internally for about a year to improve its models.

What's smart

Most AI tests focus on math, coding or difficult exam questions. LifeVal instead points to what ordinary people actually struggle with day to day: paperwork, orders and appointments. These are exactly the kinds of tasks AI companies hope their agents can take over.

The test also provides a common yardstick for progress. OpenAI's chart shows the average score of the company's own models over time:

- GPT-5.2 (December 2025): 49.3 percent - GPT-5.4 (March 2026): 73.0 percent - GPT-5.5 (April 2026): 77.4 percent - GPT-5.6 (July 2026): 80.0 percent - GPT-6 Astra (September 2026): 80.6 percent

That is a big jump from 49.3 to 73.0 percent in three months, while the most recent steps are smaller.

Cheaper or better?

It's hard to say, because LifeVal is not a new model with a price tag but a measurement method. Pricing and availability in Denmark or the EU are not stated in the sources.

There is no answer on how it compares with competitors, either. The chart shows no models from Anthropic, Google or xAI, so there is no way to tell whether OpenAI is ahead.

The blog also criticizes the lack of transparency. The task set has not been published, so outside researchers cannot verify the results themselves or test other models. The number of tasks, the grading criteria and the method itself have not been disclosed either. The blog suggests that OpenAI's internal use of the test for a year may have skewed the results in favor of the company's own models. That is the blog's assessment, not something that has been proven. The figures come from an intermediary and have not been checked against OpenAI's own announcement.

What it's good for

LifeVal is not a product you can use yourself, but a hint of where AI agents are headed. The score of about 80 percent for GPT-6 Astra suggests that OpenAI believes its agents have become markedly better at practical tasks such as shopping, health and dealing with government agencies.

For ordinary users and businesses, though, the takeaway is a cautious one: The figure is an average, and without an open method and independent tests, there is no way to know how well an agent will handle your particular task. It is worth watching whether other AI companies or researchers get access to test their own models in the same way.