OpenAI LifeVal: Can AI Finish the Errand, Not Just Explain It?

Can an AI assistant finish an errand, or does it give you a convincing explanation of how to do it? OpenAI introduced LifeVal on November 2, 2026, as an evaluation of AI help with everyday tasks. Its examples move the discussion from polished answers toward practical activities such as shopping, comparing schools and handling household information.

That distinction matters because a useful response and a useful outcome are different things. A model can summarize an enrollment process without completing enrollment. A better evaluation needs to make the work between the request and the outcome visible.

Source: OpenAI: LifeVal — November 2, 2026 ↗

OpenAI’s published chart reports 80.6% average reward for GPT-6 Astra across 169 tasks. The chart describes an unweighted measure: attempts are averaged within each task, then the 169 task means receive equal weight. The displayed percentage is reward multiplied by 100. It is not a universal completion rate for the next task you give an assistant.

The same chart places GPT-5.2 at 49.3%, GPT-5.4 at 73.0%, GPT-5.5 at 77.4% and GPT-5.6 at 80.0%. These are OpenAI’s reported evaluation results, with the metric and setup attached to the comparison.

One published school example ends with three kindergarten options organized and two preferences left for the user to decide. That is a useful handoff. It also shows why the final state matters: a prepared shortlist is different from a completed enrollment. The human decision remains part of the workflow.

Source: OpenAI: LifeVal — November 2, 2026 ↗

An evaluation score is not a receipt for your own transaction. Different tasks require different tools, permissions, checks and decisions. The announcement does not establish that every illustrated workflow is available in every ChatGPT account or that the same result will occur for your circumstances.

There is also a measurement distinction worth keeping intact. Average reward can summarize progress across an evaluation, but it cannot by itself tell you whether a particular booking was confirmed, a form was accepted or a purchase was completed. Those outcomes need their own evidence.

Before delegating, define what done means. Do you want a comparison, a draft, a submitted request or a confirmed booking? Ask the assistant to identify what it completed, what remains pending and which decisions still need you.

Think of an agent task as a chain: find information, use tools, check the result and return decisions to the person responsible. A clear handoff is useful progress. A confident paragraph that hides the unfinished steps is harder to trust.