OpenAI unveils ChatGPT agent with its own virtual computer

OpenAI just launched ChatGPT agent, giving its AI its very own "virtual computer" to browse the web, run terminal code, and finish multi-step tasks from start to finish.

The feature combines two of OpenAI's previous standalone technologies: Operator, which can visually click and navigate websites, and deep research, which excels at mining and synthesizing web data. While OpenAI originally planned to offer the agent to Pro, Plus, and Team users, massive demand on day one forced the company to delay the rollout for Plus and Team tiers, leaving it mostly in the hands of $200-a-month Pro subscribers for now.

A virtual computer with a visual browser and a terminal

Until now, AI assistants could tell you what to do, but they couldn't actually sit down at a computer and do it. OpenAI's solution is to give the agent a fully unified environment.

Inside this virtual PC, the agent can fluidly switch between multiple built-in tools. It has a visual browser for clicking through human-centric websites, a fast text browser for reading walls of text, direct API connections to apps like Gmail and GitHub, and a command-line terminal to run code. It can also produce editable presentations and spreadsheets summarizing its work.

On benchmarks, the results look remarkably strong. On Humanity's Last Exam, an expert-level evaluation, the underlying model scored 41.6 on a single attempt and 44.4 when allowed eight parallel rollouts. On the brutal FrontierMath benchmark, using the terminal pushed its accuracy to 27.4%. In real-world work simulations, it outscored Copilot in Excel by 45.5% to 20.0% on SpreadsheetBench and beat senior banking analyst baselines on complex financial modeling.

50 minutes for a lamp and phantom shopping carts

In real life, however, working with ChatGPT agent feels less like interacting with a superintelligence and more like watching a day-one intern crawl through their first shift.

During a hands-on test by The Verge, the agent was asked to find a Japanese-style lamp on Etsy under $200 and add five options to the cart. It worked for 50 grueling minutes, giving a tedious play-by-play that detailed everything from "setting up desktop" to "pressing Enter for search results."

Even weirder: the agent proudly announced it had added all five lamps to the user's Etsy cart, totaling around $825. But when checked on a real browser, the cart was completely empty. Because the agent was working inside its isolated virtual desktop without user login credentials, it had seemingly added the items to a temporary phantom cart that vanished into the ether.

When asked about the painfully slow speed, OpenAI product lead Yash Kumar and research lead Isa Fulford explained that the team prioritized hard reasoning over latency. Their view is that users shouldn't sit and stare at the screen; you are supposed to kick off a heavy chore in the background and wait for a mobile notification when it wraps up. Meanwhile, sensitive financial transfers remain firmly blocked behind safeguards like "Watch Mode," which forces users to stay on the active tab.

Why it matters

We think this launch marks the messy beginning of the true "background labor" era for AI. Bundling web interaction, code execution, and deep reasoning into an autonomous sandbox is a brilliant architectural move by OpenAI. Instead of forcing text models to guess answers, it lets the model test its work, debug with terminal commands, and inspect visual UI just like a person would.

For everyday consumers, it's clearly not ready. No one needs an AI that takes 50 minutes to browse Etsy for home decor. But for knowledge workers and financial analysts, the value proposition is real. Kicking off a task to build a three-statement LBO model or scrape competitor pricing while you sleep is genuinely useful, even if it runs slowly.

The real danger here is "action hallucination." It's one thing when a chatbot hallucinates a historical date; it's entirely another when an agent falsely claims it purchased items, submitted expenses, or booked flights. Until OpenAI fixes these phantom actions and cuts down the massive computing latency, ChatGPT agent feels more like a tantalizing $200 peek into the future than an indispensable daily tool.