SAN FRANCISCO: OpenAI has said its artificial intelligence models are increasingly capable of handling everyday tasks, including shopping and budget preparation, with performance approaching human-level capabilities.
The company has announced a new internal evaluation tool designed to assess how effectively AI technology can manage a range of routine activities, from preparing grocery lists and filling out online forms to handling more complex personal and practical tasks.
According to OpenAI, its GPT-6 Astra model achieved 98 percent accuracy in a task involving preparations for community volunteers responding to floods, while it scored 90.4 percent in developing a meal plan for a person with diabetes.
The announcement comes amid growing competition from technology companies including Meta, Anthropic and Google, all of which are developing AI agents capable of completing real-world tasks. OpenAI and its competitors are also facing increasing scrutiny over AI safety and reliability.
Through the new evaluation initiative, OpenAI sought to determine how accurately ChatGPT can perform everyday activities and understand users’ practical needs.
OpenAI researcher Jenny Yang told US media that, given the large number of people using the application, the company wanted to understand how effectively its AI chatbot could respond to human needs across different types of tasks.
She said the initiative would also help identify areas where the technology still falls short and determine how it could be improved to better serve users.
For the project, OpenAI surveyed 4,000 people and subsequently selected 100 users to describe specific tasks they would like AI to perform. The company also took steps to ensure that participants clearly understood the tasks they were proposing.
Overall, the GPT-6 Astra model achieved an average performance score of 80 percent across the everyday tasks included in the evaluation.