OpenAI releases benchmark to measure AI performance on daily life tasks

OpenAI announced a new internal evaluation tool on Monday designed to test how well its artificial intelligence models handle everyday human activities. The company stated that its latest model, GPT-6 Astra, achieved an overall score of 80% on these general performance prompts. This result represents a substantial improvement over the GPT-5.2 model, which was released in December 2025 and scored 49.3% on the same metrics.

The benchmark focuses on tasks that users commonly request from ChatGPT, such as finding a kindergarten, managing community volunteers, and creating meal plans. OpenAI reported that GPT-6 Astra scored 93.1% on a task involving locating a kindergarten after a move. The model earned a 98% score for managing community volunteers preparing for a flood and a 90.4% score for developing a meal plan for a parent recently diagnosed with diabetes.

OpenAI developed this new testing framework after determining that existing standards did not adequately measure how AI interacts with daily life. The company previously released a benchmark called GDPVal last year to evaluate work-related skills in sectors like finance and manufacturing. However, measuring everyday tasks presented different challenges because the steps involved are often broad and the results can be subjective.

To build the new test, OpenAI reviewed data from the Bureau of Labor Statistics’ American Time Use survey and the United Kingdom’s Multinational Time Use Survey. Researchers filtered out activities that AI could not assist with, such as sleeping and eating. The team then surveyed approximately 4,000 people to identify tasks users wished an AI assistant could perform. A smaller group of about 100 users provided specific task descriptions, personal context, and grading rubrics for those tasks.

Jenny Yang, a researcher at OpenAI, explained that the team required participants to demonstrate expertise in the specific area they were proposing for testing. "You had to like have evidence that you were some kind of like expert or really understood what the right answer looked like for you," Yang said. The evaluation measures skills such as navigating websites, applying context from chat history, and drawing conclusions that align with user expectations.

The release of this benchmark occurs as OpenAI faces competition from Meta, Anthropic, and Google in the development of AI agents capable of executing real-world tasks. Meta’s Muse agent launched in September and reached 10 million downloads in the United States within 40 days, according to Sensor Tower. OpenAI introduced its Dots agent in the same month, which handles long-running tasks in the background. Anthropic announced its Claude Duet agent app in October.

The industry is also addressing safety concerns following recent incidents where AI models breached testing environments. These events, which occurred during testing rather than with consumer products, have raised questions about the pace of development. There is also ongoing discussion about how to define boundaries for tasks users should delegate to AI.

Yang stated that the goal is to help users reclaim time by automating time-consuming errands. "I think there’s so many cases like this where being able to use AI just empowers you to do things that you wouldn’t have been able to do, or just makes your life easier overall," Yang said. The new tool allows OpenAI to identify areas where the technology needs improvement and to track progress over time.

OpenAI plans to use the results from this new benchmark to guide the training and optimization of its future models. The company aims to transition ChatGPT from a content generation tool to an assistant that can complete tasks on behalf of users. The evaluation provides a standardized way to assess performance in a field where previous testing was difficult to define.