The short version
The most important fact about OpenAI Operator is one that most write-ups still bury: it is no longer a standalone product. Operator launched as a research preview on February 1, 2025 (announced January 23, 2025) and was retired on August 31, 2025, with its capabilities folded into the newer ChatGPT agent. That single detail reshapes any honest verdict. If you arrived here hoping to sign up for Operator, you cannot; the answer is to look at ChatGPT agent instead. What follows is a review of what Operator actually was, how well it worked, and what its short life tells you about the wider category of browser-controlling agents.
Read as a snapshot, Operator was best understood as an early, deliberately cautious experiment in letting an AI drive a real web browser. It suited technically curious ChatGPT subscribers who wanted to watch an agent click through live websites and were comfortable with a preview-grade tool that failed on anything intricate. It was a poor fit for anyone needing reliable, unattended automation of important tasks, and it was never a general productivity tool you could trust with money or deadlines.
Who Operator was actually for
Based on how it was positioned and what it could do, Operator made sense for a narrow group:
- Existing ChatGPT Pro subscribers in the US, since that is where the preview started before its planned expansion to Plus, Team, and Enterprise tiers.
- Early adopters and builders who wanted to study agentic behavior directly, testing how a model reasons about a live web page rather than reading about it.
- People automating low-stakes, repetitive browser chores such as filling forms, placing simple online orders, or scheduling appointments, where a mistake was cheap to catch and correct.
It was not a fit for regulated workflows, high-value transactions, or anyone who needed the agent to run reliably without someone watching. The tool was explicitly a preview, and its behavior reflected that.
What it did, and the tradeoffs behind each capability
Operator was built on OpenAI's Computer-Using Agent (CUA) model, an approach trained to interact with graphical interfaces by looking at the screen and taking actions rather than only producing text. That framing matters: instead of calling clean APIs, the agent perceived and clicked a real page the way a person would. The upside is broad reach, since almost any website becomes a surface it can attempt. The downside is fragility, because pages change, layouts vary, and visual interpretation is error-prone.
Autonomous browsing of live websites
Operator navigated real sites and interacted with them directly. In practice this is useful when a task spans several pages and there is no API to automate it. The realistic expectation, though, is set by its own performance: on the WebArena benchmark for web interactions it scored 58.1 percent, and on OSWorld for OS-level tasks it reached 38.1 percent. Those numbers are meaningful progress for the time, but they are also well short of human-level reliability, which is why unattended use on anything that mattered was risky.
Form filling, ordering, and scheduling
The headline use cases were repetitive, structured chores: completing forms, placing online orders, and booking appointments. These are exactly the tasks where a competent agent saves the most tedium. The tradeoff is that the same tasks often touch payment details, personal data, or bookings you cannot easily undo, so the value depended heavily on the guardrails around them.
Human oversight and a sandboxed environment
Operator ran inside a virtual machine with limited internet connectivity and safety restrictions, and OpenAI paired it with built-in safety checks intended to prevent unauthorized sensitive actions. For sensitive or consequential steps, the design leaned on human confirmation rather than full autonomy. This is the most defensible part of the product: constraining an experimental agent, requiring sign-off before it does something irreversible, and running it in a controlled sandbox are the right instincts for a tool at this maturity. The cost is that genuine hands-off automation was never really on the table.
Where it fell short
Operator's limits were documented rather than hypothetical. It did not reach human-level accuracy, and it struggled specifically with intricate user interfaces and extended, multi-step workflows. In other words, the harder and longer the task, the more likely it was to stumble, which is the opposite of what you want from automation. Combine that with a research-preview label, a narrow initial rollout to Pro users in the US, and a lifespan of roughly seven months, and the picture is clear: this was a capability demonstration, not a dependable service. Anyone who built a routine around it would have had to rebuild once it was retired.
The broader lesson for evaluating any browser agent, including its successor, is to weight reliability over demos. A benchmark in the 50-60 percent range means roughly half of nontrivial web tasks need correction, so plan to supervise. You can browse other options in this space under AI Agents if you want to compare current, actively supported tools.
Pricing, as it stood
Operator had no standalone free plan and no separate fee of its own. It was bundled into paid ChatGPT subscriptions rather than sold on its own. The preview began with ChatGPT Pro subscribers in the US, with a stated plan to extend access to Plus, Team, and Enterprise users. The vendor did not publish a discrete price for Operator as a product, so the only cost was the underlying ChatGPT subscription.
| Aspect | Details (as documented) |
|---|---|
| Standalone free plan | None |
| Separate add-on fee | None documented; bundled with ChatGPT subscription |
| Initial access | ChatGPT Pro, US |
| Planned expansion | Plus, Team, Enterprise |
| Status | Retired August 31, 2025 |
What to use instead
Because Operator itself is retired, the most direct path is OpenAI's own successor, ChatGPT agent, which was introduced in July 2025 and likewise controls a virtual computer to perform multi-step tasks. Treat it as the continuation of the same idea with more maturity behind it. Beyond that, the browser-agent category has several actively maintained entrants; rather than name unverified competitors here, it is worth surveying the current field directly in the AI Agents category and reading through the wider tool directory before committing. For deeper background on how these agents are evaluated, the blog is a reasonable starting point.
Common questions about OpenAI Operator
Can I still use OpenAI Operator?
No. Operator was retired on August 31, 2025. Its capabilities were carried forward into ChatGPT agent, which is the tool to look at if you want a browser-controlling agent from OpenAI today.
What was Operator built on?
It ran on OpenAI's Computer-Using Agent (CUA) model, which is designed to interact with graphical interfaces by taking actions on the screen rather than only generating text. It operated inside a virtual machine with limited connectivity and safety restrictions.
How reliable was it?
It did not reach human-level accuracy. On published benchmarks it scored 58.1 percent on WebArena (web tasks) and 38.1 percent on OSWorld (OS-level tasks), and it struggled with intricate interfaces and long, multi-step workflows. It was a research preview, so supervision was expected.
Did Operator cost extra?
There was no standalone free plan and no separately documented fee. Access came through paid ChatGPT subscriptions, starting with ChatGPT Pro in the US and with planned expansion to Plus, Team, and Enterprise.
What kinds of tasks was it meant for?
Repetitive, browser-based chores such as filling out forms, placing online orders, and scheduling appointments, ideally with a person available to confirm sensitive steps.
