OpenAI Unveils ChatGPT Agent: A New ‘Agent Mode’ That Can Browse, Code, and Create Slides

Summary

OpenAI announced that ChatGPT can now act as an autonomous *agent*—using its own virtual computer to browse the web, run code, and produce editable deliverables such as slide decks and spreadsheets. The feature, called Agent Mode, is being rolled out today to Pro, Plus and Team subscribers, with broader availability planned for Enterprise and Education users in the coming weeks.

What the Agent Can Do

• Web interaction – visual and text‑based browsers let the model click, scroll, log in, and filter results on your behalf.
• Tool suite – a terminal, API connectors (e.g., Gmail, GitHub), and a visual browser are available for the model to choose the most efficient path.
• End‑to‑end workflows – you can ask the agent to “brief my calendar and upcoming client meetings,” “plan a Japanese breakfast for four and order ingredients,” or “analyze three competitors and build a slide deck.”
• User control – the agent always asks for permission before taking consequential actions; you can interrupt, take over the browser, or stop the task at any time.

Availability & Activation

• Immediate rollout – Pro users receive access by the end of the day; Plus and Team users get it over the next few days. Enterprise and Education accounts will be enabled in the coming weeks.
• How to enable – In any conversation, open the *Tools* dropdown and select Agent mode. The interface then shows a live narration of the agent’s steps.
• Message limits – Pro users have 400 messages per month; other paid tiers receive 40 messages, with extra usage purchasable via credits.

Performance Highlights (Reported by OpenAI)

• Humanity’s Last Exam – pass@1 SOTA of 41.6 (44.4 with parallel attempts).
• FrontierMath – 27.4 % accuracy, far above prior models.
• SpreadsheetBench – 45.5 % score, beating Microsoft Copilot’s 20 %.
• DSBench & BrowseComp – the agent surpasses human baselines by a large margin.

*These numbers come from OpenAI’s internal benchmarks; independent verification is pending.*

Safety Measures

• Explicit user confirmation before any action that could affect the real world (e.g., purchases, data writes).
• Watch Mode – critical tasks like sending email require active supervision.
• Prompt‑injection defenses – the model is trained to detect malicious instructions hidden in web pages and will refuse high‑risk actions such as bank transfers.
• Privacy controls – one‑click deletion of browsing data and secure “takeover mode” that never stores passwords.
• Biological‑risk classification – the agent is treated as a high‑risk AI under OpenAI’s Preparedness Framework, with extra safeguards and a bug‑bounty program.

Practical Implications

• Workplace – automate repetitive reporting, convert dashboards to editable presentations, and keep spreadsheets up‑to‑date without manual copy‑pasting.
• Personal life – plan travel itineraries, book dinner parties, or summarize your inbox with a single prompt.
• Verification – users can watch the on‑screen narration, request progress summaries, and pause the task to inspect intermediate results.

Limitations & Next Steps

• The agent is still in early stages and can make mistakes; slide‑deck generation is currently in beta and may produce rudimentary formatting.
• European Economic Area and Switzerland access is not yet enabled.
• OpenAI promises regular iterative improvements and expanded toolsets over time.

--- *For the most up‑to‑date details, see OpenAI’s official announcement page and the system card linked there.*