OpenAI published LifeVal on November 2, 2026, describing an evaluation of everyday AI-assistant tasks that it says it has used over the past year. It reports selected results, including best-recorded scores.
Why everyday tasks deserve their own checks
Our assessment: an assistant's usefulness depends on completing a whole job, not just producing a convincing answer. A shopping task, for example, can fail because the item, delivery address or final price is wrong even when the written recommendation looks good. Readers considering agents for personal or workplace administration should assess the actual outcome.
Read the score with the conditions attached
This is research, not a product rollout. We have not reproduced the results.
Before comparing benchmarks, ask what counts as success, how many attempts were allowed, which tools were available and how failures were counted. The accessible announcement did not resolve these questions.
A practical way to evaluate an assistant
- Choose a reversible task with a clear expected result, such as drafting a calendar plan without sending invitations.
- Set limits on spending, sharing and account changes before it starts.
- Check the saved result and any external action, rather than relying only on the assistant's completion message.
- Record failures as well as successes, including cases where it correctly stops for your decision.
This checklist is our suggested evaluation approach, not a test we performed or a guarantee of safe autonomous operation.