OpenAI Releases ChatGPT Agent With AI Task Execution (“Agent Mode”)

• Full Agentic Capability Now Live

  ChatGPT Agent transforms passive conversations into active workflows. It can autonomously plan events, shop, analyze documents, generate reports, manipulate spreadsheets, and more, using a virtual computer outfitted with a text browser, visual browser, terminal, and access to third-party data sources via APIs and connectors.

• New Proprietary Model Within o3 Family

  The agent operates on a new, unnamed model belonging to the o3 family, purpose-built for agentic reasoning and task execution. It was trained using reinforcement learning across a range of tools, including web interaction and programming environments, integrating OpenAI’s previous Deep Research and Operator tools into a unified system.

• Merging of Teams: Unified Development Approach

  The development team, composed of 20–35 people from OpenAI’s Deep Research and Operator divisions, jointly designed the agentic model to think holistically across tools, planning steps dynamically and adjusting its methods during task execution.

• Live Use Cases from Demo, Livestream, and Documentation

  Key demonstrations and user-facing examples include:

  • Planning a date night by referencing Google Calendar and OpenTable.
  • Ordering a destination wedding dress online.
  • Designing team mascot stickers and slide decks via Google Drive data.
  • Analyzing competitors and preparing pitch presentations.
  • Cooking assistance, like planning and buying ingredients for Japanese breakfast.
  • Briefing users on upcoming meetings based on news linked to their calendar.
  • Automating routine admin tasks, like reserving office parking weekly.

• Seamless Conversational to Action Transition

  Users can move effortlessly from a text conversation to a request for action within the same chat. ChatGPT Agent interprets context, reasons over instructions, and executes using its tools without requiring a reset or re-prompt.

• Integrated Multi-Tool Architecture

  ChatGPT Agent switches between tools fluidly:

  • Executes shell commands in its virtual terminal.
  • Navigates websites in GUI mode or via text browser for deeper content parsing.
  • Uses APIs and connectors (e.g., Gmail, Google Calendar, GitHub) to gather data.
  • Synthesizes findings into structured outputs like slides, summaries, or spreadsheets.

• Contextual Memory Across Tools

  The agent maintains state between tools and tasks, allowing it to:

  • Browse a page visually
  • Parse it via text browser
  • Process files using code
  • Present the result in a presentation—all in one continuous operation.

• Slideshow Feature in Development

  The current version supports generating slide decks with editable text, visuals, and charts. However, formatting polish and consistency in export (e.g., to PowerPoint) are still being improved. Support for uploading existing slide templates is planned.

• Latency Trade-off for Complex Tasks

  Tasks can take 15–30 minutes depending on complexity, but are designed to run asynchronously. OpenAI recommends users treat it as a background assistant that delivers output once finished, not as a real-time chat interface.

• User Control and Oversight

  Agent actions that have significant consequences, such as purchases, data edits, or email sending, require explicit user permission. It prompts users to log in securely when needed and can be paused, interrupted, or re-directed mid-task. “Watch Mode” prevents misuse during financial interactions by requiring the user to stay within the active browser tab.

• No Financial Transactions (For Now)

  ChatGPT Agent is currently restricted from executing direct financial transactions. This constraint remains in place until more thorough evaluations of safety and misuse potential are complete.

• Rollout and Access Structure

  The feature is rolling out now to Pro, Plus, and Team users via the composer’s “agent mode” dropdown or by typing /agent. Pro users receive access immediately, with 400 messages/month; Plus and Team users receive 40/month, with additional usage available via credits. Enterprise and Education tiers are slated for rollout later in summer 2025. No access yet for the European Economic Area or Switzerland.

• Work and Personal Use Applications

  • Professional: Competitive intelligence, financial modeling, document generation, presentation building, team scheduling.
  • Personal: Online shopping, cooking and meal planning, gift ordering, travel coordination, and event management.

• Benchmark Leadership

  • Humanity’s Last Exam (HLE): 41.6% pass@1, rising to 44.4% with multi-attempt rollout.
  • FrontierMath: 27.4% accuracy—leading all AI systems on expert-level mathematical problem-solving.
  • SpreadsheetBench: 45.5% accuracy—outperforming GPT-4o and Excel Copilot.
  • Investment Banking Modeling: Surpassed human baselines on three-statement models and LBO tasks.
  • BrowseComp: 68.9%—the highest known score, outperforming the Deep Research baseline by 17.4 points.

• Safety, Privacy & Risk Mitigation

  • Prompt Injection Resistance: Trained to detect and neutralize malicious prompts embedded in web content.
  • Explicit Confirmation for Critical Actions: Mandatory for all irreversible steps.
  • Data Controls: Browsing data can be deleted in one click. Takeover mode does not expose passwords or user credentials.
  • High-Risk Classification: Categorized under OpenAI’s High Biological and Chemical Capabilities classification, even without evidence of misuse, to trigger maximum safety protocols.
  • Collaboration on Security: Ongoing partnerships with security institutes, government agencies, and academic red teams. A bug bounty program is live to catch real-world risks.

• AI Agent Trend & Industry Context

  ChatGPT Agent is a key competitor in the growing AI agent race. Since Klarna replaced ~700 support roles with its AI system in 2024, major tech companies have invested heavily in agentic infrastructure. Meta’s multibillion-dollar investment in compute and its $200M hiring of an Apple AI lead underscore the strategic priority. OpenAI’s agent joins similar tools like Claude’s “Computer Use” and Google’s ongoing agentic projects.

• A Tangible Step Toward AGI

  OpenAI CEO Sam Altman described the release as a “feel the AGI moment,” highlighting the agent’s ability to think, plan, and act within a digital environment. Though still experimental, it signals a future where AI can act as an autonomous assistant across work and life contexts.

Why This Matters:

ChatGPT Agent is more than a chatbot, it’s a doer. With a dynamic, modular architecture, best-in-class benchmarks, and a focus on user safety, it reflects the next phase of AI evolution. This release moves AI closer to real-world utility and toward the long-term vision of AGI, an intelligent system that not only understands but acts meaningfully on human intent. As enterprise, infrastructure, and consumer applications align, OpenAI’s agent sets a high bar for the future of productivity.