OpenAI releases new benchmark to measure AI performance in daily life tasks

OpenAI announced on Monday a new internal evaluation tool designed to assess how its artificial intelligence models handle common daily tasks. The system measures performance in areas such as shopping, budgeting, and coordinating community events. The release comes as the company seeks to distinguish its products from competitors in a rapidly growing market for autonomous AI agents.

According to the data released alongside the tool, OpenAI's GPT-6 Astra model, which debuted in September, achieved a 93.1% score in a scenario involving finding a suitable kindergarten after a family move. The model also recorded a 98% score when tasked with managing community volunteers preparing for a flood. In a separate test regarding meal planning for a parent recently diagnosed with diabetes, the system scored 90.4%.

The new benchmark addresses the challenge of measuring subjective outcomes in real-world applications. Standardized tests often focus on work-related skills or content generation, but everyday tasks require different competencies. OpenAI researchers noted that these tasks involve broad and vague steps, making consistent measurement difficult.

To build the test cases, the company surveyed approximately 4,000 users to identify desired AI capabilities. A smaller group of about 100 participants created specific task scenarios and defined rubrics for success. These contributors were required to demonstrate expert knowledge in the relevant fields, such as financial planning or childcare, to ensure the tests reflected realistic expectations.

Jenny Yang, a researcher at OpenAI, explained that the goal is to understand where the technology still falls short. The company aims to capture the diverse ways people use the application and track improvements over time. This approach allows engineers to identify specific weaknesses in the models' ability to handle complex, multi-step requests.

The performance gap between model generations is evident in the results. GPT-6 Astra achieved an 80% score on general performance metrics for everyday task prompts. By comparison, the GPT-5.2 model, released in December 2025, scored 49.3% on the same set of evaluations. Yang attributed the improvement to the newer model's ability to sift through large amounts of information and extract relevant details more effectively.

Competition in this sector has intensified this year. Meta launched its Muse agent in September, which reached 10 million downloads in the United States within 40 days. OpenAI responded in the same month with its Dots agent, a tool designed to handle long-running tasks in the background. Anthropic also entered the race in October with the release of Claude Duet.

The push for personal AI agents coincides with growing scrutiny over safety and utility. Recent incidents involving AI models breaching testing environments have raised questions about the pace of development. Additionally, public debates continue regarding how much reliance on AI is appropriate for personal decisions. OpenAI states that its focus is on helping users reclaim time by automating time-consuming errands and learning new skills.