OpenAI introduces LifeVal, an evaluation of everyday tasks for ChatGPT

OpenAI introduced LifeVal, an evaluation of everyday tasks it built to measure how well its models help people with practical needs. OpenAI says it used LifeVal to track and drive improvements over the past year, with the goal of moving ChatGPT from a chat tool toward an agentic personal assistant that takes actions.

Details

• Task selection: Tasks come from the American Time Use Survey, the Multinational Time Use Study, and a new demographically representative survey of 3,995 respondents. Areas covered are caregiving, health, education, household administration, personal finances, shopping, and government services.
• Sample scores: Kindergarten enrollment after a move, 93.1%; managing flood-preparation volunteers, 98.0%; finding a family home, 100%; setting up a remote game-watch with a grandparent, 96.3%; meal planning after a parent diagnosed with diabetes, 90.4%.
• Overall result: OpenAI says its latest models score very highly on many LifeVal tasks and credits training improvements in browser use, computer use, and connectors. No overall score is given.
• Undisclosed: Total number of tasks, named model comparisons, and authors are not stated.

What happened next

The sample scores are close to ceiling, so the main takeaway is the task-selection method: grounding an agent benchmark in how people actually spend their time. Without the full task list or model-by-model results, the numbers cannot yet be compared independently.