OpenAI has launched Operator, an AI agent capable of performing online tasks like booking concert tickets and grocery shopping. This web app, powered by the new Computer-Using Agent (CUA) model, is built on top of OpenAI’s multimodal large language model, GPT-4o. Currently, Operator is available to US users subscribed to ChatGPT Pro, a $200-a-month service. OpenAI plans to expand access in the future. Operator outperforms similar tools from competitors, scoring 87% on the WebVoyager benchmark for browser tasks, compared to 83.5% for Google DeepMind’s Mariner and 56% for Anthropic’s Computer Use. On the OSWorld benchmark, which tests tasks like merging PDFs, CUA scores 38.1%, while Computer Use scores 22.0%. Humans, for comparison, score 72.4%. Operator runs on remote servers, allowing it to handle multiple tasks simultaneously, and collaborates with businesses like OpenTable and Instacart to streamline user interactions.
Source: www.technologyreview.com















