OpenAI Unveils ChatGPT Agent: A New Era of AI Task Automation

OpenAI has launched a ChatGPT Agent that enables the AI to autonomously handle complex tasks using its own virtual computer. This is a huge leap forward, blending research and action into a unified agentic system designed to think and act on behalf of users.

Revolutionary Capabilities and Task Automation

The ChatGPT Agent can now tackle sophisticated requests such as "review my calendar for client meeting updates based on recent news," "organize and purchase ingredients for a Japanese breakfast for four," or "analyze competitors and build a slide deck." It navigates websites, filters data, prompts secure logins when needed, executes code, and delivers editable outputs like slideshows and spreadsheets. This capability integrates three key strengths from their earlier innovations: Operator's web interaction skills, deep research's synthesis abilities, and ChatGPT's conversational fluency.

How the Virtual Computer System Works

Operating on its own virtual computer, the agent seamlessly shifts between reasoning and action, managing intricate workflows based on user instructions. Users remain in control, with the agent requesting permission for significant actions and allowing interruptions or manual browser takeovers at any time. Starting today, Pro, Plus, and Team users can activate these agentic features directly through the tools dropdown in the composer by selecting "agent mode" during any conversation.

Advanced Tool Integration and API Access

The launch is just the beginning, with OpenAI planning regular improvements to enhance its capabilities and accessibility. The agent is equipped with a suite of tools, including a visual browser for web interaction, a text-based browser for reasoning, a terminal, and direct API access. It also leverages ChatGPT connectors, allowing integration with apps like Gmail and GitHub to pull relevant data into responses. For example, it can access calendar information via APIs, reason over large text volumes, and visually engage with human-designed websites, adapting its approach for efficiency.

Collaborative Workflows and User Control

This setup supports iterative, collaborative workflows, where users can interrupt to refine instructions or change tasks, with the agent resuming without losing context. It may proactively seek clarification to align with user goals and can be paused for progress updates or stopped for partial results. The agent excels in diverse scenarios, from automating financial model updates at work to planning dinner parties personally, showcasing its versatility.

Performance Benchmarks and Results

Their performance shines on benchmarks: the model powering ChatGPT Agent achieves a 41.6% pass on Humanity's Last Exam, rising to 44.4% with multiple attempts, and a 27.4% accuracy on FrontierMath with tools, outpacing prior models. On complex tasks like competitive analyses or investment banking models, it matches or exceeds human performance in about half the cases, leads with 87.9% on DSBench data analysis, and scores 45.5% on SpreadsheetBench with .xlsx access. It also sets new records on BrowseComp (68.9%) and WebArena (78.2%).

Safety Measures and Risk Mitigation

However, this expanded capability introduces new risks, as the agent can act on the web and access sensitive data via connectors or logins. OpenAI has implemented safeguards, including user confirmation for consequential actions, active supervision for tasks like emailing, and refusals for high-risk operations like bank transfers. Privacy controls allow data deletion and secure browser takeovers, while enhanced mitigations address prompt injection risks, which are malicious web instructions that could manipulate the agent. Given its biological and chemical potential, they've applied their strictest safety protocols, collaborating with biosecurity experts to mitigate dual-use concerns.