OpenAI’s LifeVal tests AI on everyday tasks, not just knowledge

OpenAI announced LifeVal on November 2, 2026, an evaluation designed to measure how well AI models handle everyday tasks rather than only academic knowledge. In its November 2 post, OpenAI reports five example task scores ranging from 90.4% to 100%.

These are OpenAI's own results on its own benchmark, not independent measurements. As of that post, OpenAI does not name the models behind the percentages or explain their grading method.

#What does LifeVal measure?

As of November 2, 2026, OpenAI describes LifeVal as an evaluation of tasks in areas where users wanted the most help from AI and personal agents. It says it used the evaluation to track and drive improvements during the preceding year.

The task areas include caregiving, health, education, household administration, personal finances, shopping and access to government services. These are evaluation categories, not evidence that a model can reliably complete every task in those areas.

OpenAI explains its aim this way: "To be a truly helpful personal assistant, ChatGPT should be able to take actions that actually get things done."

#How were the tasks selected?

OpenAI's November 2 account says it started with the American Time Use Survey and the Multinational Time Use Study, which track how people spend time in the US and internationally. It then added a new, demographically representative survey of 3,995 respondents to understand where people would most value AI assistance.

India Today reported on November 4, 2026 that the multinational study combines more than one million diary days from over 100 national surveys, covering 30 countries over 55 years. Those scale figures are India Today's description of the underlying study, not a count of LifeVal tasks.

India Today's November 4 report describes the American survey as tracking time spent on paid work, childcare, volunteering and socializing. It frames LifeVal around whether an AI finishes work and meets a person's requirements. OpenAI remains the source for the benchmark's design and reported results.

#What scores did OpenAI report?

The following titles and percentages are from OpenAI's November 2, 2026 post. OpenAI labels the performers only as "our current models" and separately refers to "Our latest models". No named model should be inferred from these results.

These are five examples, not a disclosed total task count or an overall benchmark score. The percentages should not be treated as independently validated success rates for real-world users.

#How does this connect to agent capabilities?

As of its November 2 post, OpenAI says it had trained models to become faster, more capable and more aligned on browser use, computer use and connectors tasks during the preceding year. It argues that enabling those capabilities in ChatGPT expands the kinds of tasks the model can help with, moving it beyond chat toward an agentic personal assistant.

For separate coverage of tools and access, see our reports on OpenAI's computer-use development and ChatGPT and Codex's shared usage allowance. Neither report identifies which models produced the LifeVal scores.

Our analysis: Evaluating a completed workflow asks a different question from testing what a model knows. LifeVal's everyday-task framing is useful for that distinction. But the five examples leave important gaps: without named models, a total task count and a grading method, readers cannot use the percentages to make like-for-like comparisons or judge reliability across all the listed categories.

#What is unclear?

• Model identity: OpenAI does not name the models behind the November 2 results.
• Coverage: the total number of LifeVal tasks is not stated.
• Scoring: the method used to grade the percentages is not explained.
• Reproducibility: OpenAI does not state whether the tasks or data are publicly released for others to run.