OpenAI Operator
AI agent that uses its own browser to navigate websites, complete tasks, and take actions on your behalf.
What it is
OpenAI Operator is an AI agent that browses the web and completes tasks autonomously on your behalf. It was launched in January 2025 as a research preview for ChatGPT Pro users and has since been fully integrated into ChatGPT as "ChatGPT Agent" mode, available to Plus, Pro, Business, Enterprise, and Edu subscribers. The standalone Operator website (operator.chatgpt.com) is no longer accessible; all functionality now lives inside ChatGPT agent mode. Operator is powered by GPT-5.4, which includes native computer use capabilities, replacing the earlier CUA model that combined GPT-4o's vision with reinforcement learning. The agent processes screenshots as raw pixels and takes actions through a virtual mouse and keyboard, meaning it can interact with any website exactly as a human would. On March 11, 2026, OpenAI launched the Responses API computer environment, enabling developers to build their own computer-using agents programmatically. CUA (Computer Use Agent) is also available directly within ChatGPT agent mode. The CUA model is available via API at $1.50 per 1M input tokens and $6 per 1M output tokens. Operator includes a three layer safety architecture: user control (takeover mode for sensitive actions), data privacy protections, and adversarial defenses including a monitor model with 99% recall on prompt injection attempts. The system supports partner integrations with services like DoorDash, Instacart, Uber, OpenTable, and Priceline. Agent mode is available in all ChatGPT supported countries and territories, including the EU/EEA and Switzerland.
Who it is for
Power users wanting AI to handle web tasks like shopping, booking, research, and form filling without manual navigation
Developers building computer-using agents via the CUA API for custom browser automation and testing workflows
Busy professionals automating repetitive web workflows that currently require manual clicking through multiple pages
Teams using ChatGPT Pro who want maximum AI agent capabilities for complex, multi step web tasks
Anyone who spends significant time on routine web tasks (ordering food, booking restaurants, comparison shopping) and wants to delegate them to an AI
Pros and cons
Pros
- +Can complete real world tasks on actual websites, not just answer questions about them; it navigates, clicks, fills forms, and completes purchases
- +Partner integrations with DoorDash, Instacart, Uber, OpenTable, and others provide optimized, reliable task completion for common services
- +Three layer safety architecture prevents unauthorized actions: takeover mode for sensitive steps, data privacy protections, and adversarial defenses
- +Now powered by GPT-5.4 with native computer use, delivering improved accuracy and reasoning compared to the earlier GPT-4o based CUA model
- +CUA model available via API, including the new Responses API computer environment (March 2026), allows developers to build custom computer-using agents for specialized automation workflows
- +Background execution lets complex tasks (15 to 30 minutes) run while you continue working, with notifications when complete or when input is needed
Cons
- −No free tier; requires at minimum a ChatGPT Plus subscription ($20/month) to access agent mode, and agent message limits are relatively low (40/month on Plus, 400/month on Pro)
- −Reliability varies by website complexity; simple tasks succeed often, but complex multi step workflows on poorly designed sites can fail
- −Scored 38% on the OSWorld benchmark, showing significant limitations on complex desktop UI tasks and non standard web interfaces
- −Complex tasks can be slow, taking 15 to 30 minutes to complete workflows that a human might finish in 5 minutes with direct navigation
- −Financial transactions are restricted; the agent requires explicit user confirmation for purchases and cannot autonomously spend money
- −Prompt sensitivity affects success rates; slight changes in how you describe a task can lead to different outcomes or failures
Pricing
- Agent mode access in ChatGPT
- Web browsing and task completion
- Partner integrations
- Takeover mode for sensitive actions
- All Plus features
- 5x or 20x more usage than Plus
- 400 agent messages per month
- $1.50 per 1M input tokens
- $6 per 1M output tokens
- Full CUA model access
Prices change; check the official pricing →