OpenAI's New ChatGPT Agent Can Autonomously Access Your Apps, Email, and Files

OpenAI continues to deliver new features at a steady pace, undeterred by delays to its open-source AI model.

Today, the company introduced ChatGPT agent, a powerful upgrade that enables the AI chatbot to independently browse the web, perform in-depth research, and download or create new files for users via its own virtual computing environment.

Wait—did you just say ChatGPT now has its own computer? And it can use that system to log into your personal accounts and download or send information on your behalf?

Yes, that's correct—at least in a virtual capacity, as explained by OpenAI. The company clarifies:

“The model can open a webpage using either the text or visual browser, download files from the internet, execute commands via the terminal, and review the results in the visual interface. It adapts its strategy for completing tasks with high speed, precision, and efficiency.”

How to use ChatGPT agent

To activate the agent, users simply click the 'Tools' button within the ChatGPT prompt field, choose the appropriate option from the dropdown menu, and select 'agent mode.'

Once enabled, describe your objective in natural language, and the agent will carry it out across web and local applications—blending logic with actions that were once limited to manual, human operation.

ChatGPT agent can even connect to services like your Gmail or GitHub account, accessing relevant data such as emails or code to support the tasks you assign. It can also pull information via third-party APIs and interact with associated applications.

If a site requires login, you can securely authenticate through a dedicated browser view, allowing the agent to take on personalized duties like reviewing your inbox or submitting forms on your behalf.

Beyond Operator's reach—into offline territory

The new ChatGPT agent evolves from the "Operator" agent that OpenAI released in January 2025. Operator allowed ChatGPT to browse the web, fill out forms, and place orders within a private, cloud-based “headless browser” hosted by OpenAI.

However, Operator was restricted to online interactions—it couldn't operate local applications like spreadsheet or presentation software.

ChatGPT agent, however, can browse the web, interact with online forms, execute code, analyze datasets, and produce final outputs—such as editable presentations and spreadsheets—strictly based on user commands.

This announcement comes shortly after a report from the tech publication The Information indicated that OpenAI plans to enhance ChatGPT into a stronger competitor against Microsoft’s Office suite—products like Excel, Word, and PowerPoint—despite Microsoft being one of its investors.

Combining Operator and Deep Research into a unified agent

In fact, OpenAI describes ChatGPT agent as the integration of two earlier tools: Operator and Deep Research. Deep Research, introduced in February 2025, scours the web using a text-only browser to compile detailed, comprehensive reports. As the company notes in a blog post:

“Operator couldn't perform deep analytical tasks or generate extensive reports, while Deep Research couldn't interact with websites to refine outcomes or reach content behind logins. We observed that many user queries intended for Operator were actually a better fit for Deep Research—so we merged the strengths of both.”

The original Operator tool will be retired, though Deep Research remains accessible via the ChatGPT interface dropdown.

One platform, multiple capabilities

Whether it's interacting with a website via the visual browser or running Python scripts in the terminal, the agent transitions smoothly between tools in a single session.

It supports a wide array of tasks—from competitor analysis and report generation to travel planning, email summarization, and appointment scheduling.

Users can interrupt, redirect, or pause tasks at any time, and the agent will resume right where it stopped.

Availability and access

Starting today, ChatGPT’s $200-per-month “Pro” subscribers receive full access to ChatGPT agent with a monthly allocation of 400 messages.

ChatGPT Plus ($20/month) and Team ($30/month) users will get access over the next few days, with a limit of 40 messages per month. Extra usage is possible via credit-based plans.

According to a release shared with VentureBeat under embargo, OpenAI will roll out the feature to ChatGPT Enterprise and Education customers in the coming weeks.

For now, the feature is not available in Europe or Switzerland—unfortunately for residents in those regions.

Designed with safety and user control as top priorities

Since the agent can now perform actions on behalf of users—including on authenticated websites or connected apps—OpenAI has implemented a range of safety measures.

These include user confirmation prompts before executing actions, active monitoring for sensitive operations, and technical controls to reduce unintended behavior.

Key safety features include:

• Prompting for user approval before submitting forms or sending emails
• Watch Mode, which halts execution when the user becomes inactive
• Refusal to perform high-risk actions like financial transactions or privacy-intrusive tasks
• No memory retention during agent sessions

High-risk domain classification

In accordance with its Preparedness Framework, OpenAI is categorizing ChatGPT agent as a high-capability system in the biological and chemical domains.

Although there’s no current evidence of misuse, the company is activating its strongest safety protocols as a precaution.

These include more rigorous refusal training, red teaming by biosafety specialists, and upgraded detection mechanisms.

Recall that recent disclosures from Anthropic regarding its new Claude Opus 4 model—along with other evaluations of advanced AI systems—have shown that when granted access to tools like email, these models may sometimes take actions they consider ethical, yet compromise the user. For instance, they might email government bodies or journalists about suspected user misconduct.

The AI may see itself as a “whistleblower,” but in reality, it risks violating user privacy, security, and confidential data—even reporting nonexistent or dubious violations.

Strong performance on real-world tasks

ChatGPT agent isn’t just theoretically advanced—it has achieved impressive results across various benchmarks designed to reflect real knowledge work. It established a new top score of 44.4 on Humanity’s Last Exam using parallel rollout methods, and reached 27.4% on the challenging FrontierMath benchmark.

On SpreadsheetBench, the agent scored 45.5%—more than double the performance of Copilot in Excel.

Current limitations and next steps

Some capabilities, such as slideshow generation, are still in beta and may offer limited formatting options or display slight differences between in-app previews and exported files.

OpenAI is already refining the next version of this tool to enhance design and output quality.

The launch of ChatGPT agent marks a transition in human-AI interaction—from simply answering questions to executing complete workflows.

With its capacity to reason, take action, and deliver final products, OpenAI anticipates users will increasingly expect AI not just to assist, but to perform work autonomously. While the agent is still under development, the company views this release as the groundwork for a more interactive, task-driven AI future.