LifeVal Operational

LifeVal is an evaluation benchmark developed by OpenAI to measure AI's ability to help people complete everyday tasks. It was built from the American Time Use Survey, cross-national time-use research, and a representative survey of 3995 people. Its tasks cover areas such as childcare, health, education, household administration, personal finance, shopping, and accessing government services, emphasizing whether a model can actually "get things done" rather than merely offer advice. Official data shows that LifeVal has been used over the past year to measure and improve model performance. Based on the unweighted average reward, the average score increased from 49.3% for GPT-5.2 to 80.6% for GPT-6 Astra.

Measures everyday task ability: Evaluates model performance in real-life scenarios involving childcare, health, education, finance, shopping, and government services.
Built around real user needs: Draws on time-use surveys and a representative survey of 3995 people to select tasks where users most need AI assistance.
Evaluates end-to-end task completion: Focuses on whether a model can "get things done," rather than merely provide chat-style answers or suggestions.
Tracks model progress: Enables longitudinal comparison. Official data shows scores rising from 49.3% for GPT-5.2 to 80.6% for GPT-6 Astra.
Drives Agent capability improvements: Supports training and optimization for browser use, computer operation, connector tool calls, and other capabilities.
Goes beyond academic evaluation: Shifts from testing "knowledge" to measuring AI's value in helping ordinary people handle practical matters.

Task sources: LifeVal tasks come from three types of data: the American Time Use Survey (ATUS), the Multinational Time Use Study (MTUS), and an OpenAI-created population-representative survey covering 3,995 respondents. Everyday activities that people actually spend time on are extracted from time-use surveys. Questionnaires then ask what help people most want from AI for each task, identifying high-value tasks in areas such as caregiving, health, education, household administration, finance, shopping, and government services.
Evaluation format: Each task simulates a real user-help scenario and requires the model to complete the full process: understand constrained needs, collect and compare information, make trade-offs, and deliver a usable result. The evaluation focuses not on the quality of the response itself, but on whether "the task was actually completed," which fundamentally distinguishes it from answer-based academic benchmarks.
Scoring mechanism: OpenAI uses the unweighted average reward as the core metric, averaging scores assigned to the completion result of each task. This implies the existence of some evaluator that assigns a reward value based on the completeness of the model's deliverable. The design supports large-scale automated evaluation and longitudinal comparisons across versions, while the specific composition of the reward function is currently opaque.
Capability requirements: High-scoring LifeVal tasks require models to actually use browsers, computer interfaces, and third-party connectors rather than answer solely from knowledge encoded in their parameters. The benchmark essentially evaluates three dimensions at once: planning and reasoning, tool use, and self-correction during multistep execution. A model cannot achieve high rewards without all three.

Tasks originate from real needs: Built from time-use surveys and a representative survey of 3,995 people, directly aligning with scenarios where ordinary people most want AI's help.
Measures "getting things done," not "getting answers right": Evaluates end-to-end task completion rather than knowledge-question accuracy.
Covers core areas of life: Includes high-value everyday scenarios such as childcare, health, education, finance, shopping, and government services.
Advances Agent capabilities: Evaluates combined abilities including planning, browser operation, computer use, and connector calls.
Supports longitudinal progress tracking: Uses a consistent average-reward metric to measure model iterations, with official results showing a clear improvement curve from GPT-5.2 to GPT-6 Astra.
Moves beyond academic evaluation: Shifts the focus from "how intelligent is the model" to "how much practical help does it provide in users' lives?"
Closely aligned with product strategy: Used as an internal training regression test, directly supporting ChatGPT's transition from a chat tool to a personal Agent.

Project URL: https://openai.com/index/lifeval/

Model capability regression testing: OpenAI uses it internally to track everyday task performance from GPT-5.2 to GPT-6 Astra and ensure that new versions do not regress.
Guiding Agent feature iteration: Uses weaknesses revealed by the benchmark to target improvements in browser operation, computer use, and connector calls.
Evaluating new releases: Serves as a "daily usefulness" scorecard for model releases, demonstrating to users that a new version can get more things done than its predecessor.
Prioritizing product features: Uses task-completion data to determine which life scenarios, such as childcare and government services, most need research and development resources.
Reference for industry evaluation: Other AI companies can draw on its user-research-driven task-design methodology to build similar everyday-task evaluations for Chinese users.