The benchmark builds on OpenAI's earlier GDPVal evaluation, which measures AI performance on professional assignments. LifeVal instead focuses on personal responsibilities, including caregiving, education, health, shopping, financial planning, and household management.
To identify which activities people most wanted help with, OpenAI examined two existing research datasets: the American Time Use Survey and the Multinational Time Use Study. These provided information about how people allocate their time across different activities and demographic groups.
However, frequently performed activities do not necessarily represent useful opportunities for AI assistance. OpenAI therefore conducted an additional survey of 3,995 demographically representative respondents to determine which responsibilities people would prefer to delegate to a personal assistant.
The company then worked with approximately 100 users to develop evaluation tasks, according to CNN. Participants described real-world requests, supplied relevant personal circumstances, and helped establish criteria for determining whether an AI-generated result would be satisfactory.
OpenAI also required contributors to demonstrate experience with the subjects they were evaluating. For instance, someone developing a financial-planning task needed an established history of working with a financial adviser.
This approach allows LifeVal to assess more than whether an AI response is technically correct. A successful recommendation must also account for practical constraints, individual preferences, and the specific outcome the user wants.
OpenAI's published examples illustrate the complexity of those requirements.
One assignment asks an assistant to identify a suitable family home in Luzerne County while considering affordability, a separate living area, and commuting distance after school drop-off. Another involves helping a parent adjust to a diabetes diagnosis through meal planning, reminders, and questions for healthcare professionals.
Other scenarios include comparing kindergarten options after relocating, organizing volunteers ahead of potential flooding, and arranging a way for family members in different locations to watch a sporting event together.
The examples also demonstrate that completing a task does not always mean making the final decision independently. In the kindergarten scenario, the assistant prepares a shortlist but leaves unresolved preferences for the user to review. Similarly, the diabetes-related example presents a plan for consideration rather than automatically sharing it.
OpenAI reports that its newer models have made significant progress on the benchmark.
According to the company's published results, GPT-5.2 achieved an average LifeVal reward of 49.3%, compared with 73.0% for GPT-5.4, 77.4% for GPT-5.5, and 80.0% for GPT-5.6. GPT-6 Astra recorded the highest reported score at 80.6%.
The results indicate a substantial improvement across successive models, although the difference between GPT-5.6 and GPT-6 Astra is relatively small.
OpenAI also measured the speed and estimated cost of completing these tasks. The company says GPT-6 Astra is approximately 23% faster than GPT-5.6, while GPT-5.6 delivers comparable overall performance at roughly half the estimated API cost.
These measurements reflect performance within OpenAI's evaluation framework rather than a guarantee that every real-world task will be completed successfully.
The company attributes its progress partly to improvements in browser navigation, computer interaction, and connectors that allow ChatGPT to work with external services. These capabilities enable assistants to perform actions across multiple steps instead of limiting their assistance to written instructions.
OpenAI researcher Jenny Yang told CNN that LifeVal is intended to help the company identify both improvements and remaining weaknesses in its models.
The benchmark also addresses a challenge that is less prominent in conventional AI testing: personal tasks often have several acceptable outcomes, and the best result can depend heavily on information that is specific to one individual.
By incorporating user-provided context and evaluation criteria, OpenAI is attempting to measure whether its models can satisfy those requirements rather than simply generate plausible responses.
The company says more than one billion people use ChatGPT each week, making everyday assistance an important focus of its development efforts.
LifeVal provides a more structured way to assess that ambition. Its results show progress in handling complex personal workflows, while the published examples reinforce that some decisions still depend on user preferences and approval.
For OpenAI, the benchmark establishes a measure of progress toward making ChatGPT an assistant that can carry out practical responsibilities, rather than one that primarily explains how users can complete them.
This analysis is based on reporting from OpenAI.
Image courtesy of OpenAI.
This article was generated with AI assistance and reviewed for accuracy and quality.
