ChatGPT Agent Review 2026: Pricing, Features and Verdict

What it is

OpenAI's agent has gone by three names. It started as Operator, a separate site that OpenAI has since shut down. OpenAI then merged Operator into ChatGPT as agent mode: pick it from the tools menu or type /agent, describe the task, and ChatGPT works in a virtual browser, pausing for confirmation when needed. On July 9, 2026, OpenAI launched ChatGPT Work, "an agent for longer, more involved tasks," along with a new ChatGPT desktop app that combines Chat, Work and Codex.

In practice, when people search for ChatGPT agent today, they mean Work: give it a goal, follow its progress, answer its questions, change direction and approve important actions. On desktop it can use local files and desktop apps with your permission. The built in browser gathers information from sites and works with web based tools. OpenAI also retired its Atlas browser on August 9, 2026, moving those browser agent ideas here.

Pricing explained

Plan	Price	Agent access
Free	$0	Limited ChatGPT Work in the desktop app
Go	$8 per month	Limited Work in the desktop app; plan may include ads
Plus	$20 per month	Work on desktop, web and mobile; scheduled and event triggered tasks; limited GPT-6 Astra in Work
Pro	$100, $200 or $500 per month	Longer Work sessions, Dot always on agent; Pro 100 and 200 use their full allowance for Astra; Pro 500 adds Ultrafast
Business, Enterprise, Edu	See ChatGPT pricing	Work with admin controls and no training on business data by default

How usage is counted. Work "follows the same usage structure as Codex," so Work and Codex draw from one allowance. OpenAI's Codex pricing page gives estimates, not fixed limits: on Plus and Standard Business, 15 to 160 local messages per five hours on GPT-6.1 Sol and 350 to 3,000 on GPT-6 Luna. Pro plans currently have no five hour limit. Weekly limits may also apply, and Fast mode uses your allowance at 2.5x the Standard rate.

The older agent mode cap. OpenAI's agent mode help article lists monthly limits of 40 agent messages on Plus and 400 on Pro, counting only the first user request of each task. Business and Enterprise on flexible pricing pay 30 credits per agent message.

Going over. Plus and Pro users can buy credits after hitting limits, which also cover ChatGPT for Word, Excel and PowerPoint. OpenAI shows credit prices in the app rather than on the public pages we read.

Value check. At $20 per month Plus is the cheapest way into a full agent from a major lab. Pro's $100 to $500 tiers are for people who run Work or Codex for hours daily. See our ChatGPT agent pricing breakdown.

Key features

Plan mode

Work can gather context, ask questions and write a step by step plan before it starts. You edit or approve the plan, then it runs.

Finished deliverables

Work produces polished documents, presentations, spreadsheets, reports, analyses and Sites, using your templates and preferred formats. Sites turns data into interactive dashboards, trackers or prototypes you can share.

Built in browser

The desktop app's browser supports multiple tabs and richer agent workflows. Since August 25, 2026, Work's browser on web and mobile can also complete tasks on sites that need sign in for Plus and Pro: ChatGPT shows the login screen, you enter credentials (password managers work), and the session can stay signed in for later tasks. You can delete browser data per site under Settings, Cloud browser.

Plugins and connected apps

More than 1,400 plugins pull context from your tools. OpenAI's Library includes Google Drive, Box, Dropbox and SharePoint. The ChatGPT browser extension now works in Chrome, Edge, Brave, Opera and Vivaldi, so Work and Codex can use your open tabs or complete a browser task from the desktop app.

Scheduled and event triggered tasks

Tasks can run once, repeat on a schedule, or start when something happens: a new Gmail message, a new Slack message or GitHub pull request activity. Event triggers are available on Plus and higher, not on Free or Go.

Voice

ChatGPT Voice works inside Work on web and mobile and in the desktop app, so you can start a task by speaking and let it continue in text.

Dot

Pro plans add Dot, "your always-on agent," which OpenAI says has its own cloud computer and keeps making progress between conversations.

What users say

ChatGPT's iOS app holds 4.8 on the App Store from about 11.4 million ratings, the largest review base of any AI product. Agent features do not have their own rating. Common praise in public discussion is the breadth of output (files, decks, Sites) and how much it does inside one subscription. Common complaints are opaque usage limits, tasks that stall on complex sites, and having to check its work.

Privacy and security

Screenshots and takeover. The agent sees pages through screenshots of its virtual browser. When you take over to type a password, screenshots are not captured. Chats, agent browsing history and screenshots stay in your history until you delete the chat, and deleted items leave OpenAI's systems within 90 days.

Training. Business, Enterprise and Edu: no training on business data by default, including agent sessions. Plus and Pro: data, including screenshots, may be used to improve models if you have opted in; turn off "Improve the model for everyone" under Data Controls to stop it.

Human review. A limited number of authorized OpenAI staff and trusted providers may access agent content to investigate abuse or security incidents, provide support, handle legal matters, or improve models unless you opted out.

Admin controls. Enterprise owners get a workspace toggle for agent mode (off by default), role based access, app allow lists, Compliance API logging of agent conversations and a domain blocklist.

Prompt injection. OpenAI's help article gives a concrete example: an agent looking for a group dinner spot reads a malicious comment telling it to fetch a password reset code from Gmail and send it to an attacker's site. Its safeguards include confirmations for high impact actions, prompt injection monitoring and watch mode on certain sites, but OpenAI says they "don't eliminate all risks." Lockdown Mode, under Settings, Security, restricts live browsing, deep research, agent mode and file downloads for people who want to shut that door. OpenAI's own advice: avoid prompts like "check my email and handle everything," enable only the apps a task needs, and stop a task if anything looks off.

Who it is for

• ChatGPT Plus subscribers. The agent is already in your plan.
• Office workers who need output. Reports, decks and spreadsheets from your own files and apps.
• People who automate recurring work. Scheduled and event triggered tasks from Gmail, Slack and GitHub.
• Teams on ChatGPT Business or Enterprise. Admin controls and no training by default.

Who should skip it

• People who want predictable task counts. OpenAI publishes ranges. Gemini in Chrome states 20 auto browse tasks a day on Pro.
• Heavy parallel research. Manus runs 20 concurrent tasks and Wide Research on its $20 plan.
• Anyone who wants the agent inside their main browser. Comet puts it in every tab for free.

Alternatives

• Manus: parallel agents and app building, billed in credits.
• Perplexity: Computer agent and cited research from $17 per month billed annually.
• Claude: Claude in Chrome and Claude Code on Pro at $20 per month.
• Genspark: Super Agent with a weekly Free Quota for office tasks.

See all ChatGPT agent alternatives, compare ChatGPT agent vs Manus, or read the best AI agents.

Verdict

ChatGPT agent, in its ChatGPT Work form, is the most complete general purpose agent you can buy for $20 a month: it plans, browses, signs in where you allow it, works across 1,400+ plugins and returns finished work. Its weak spots are metering you can only estimate and the security tradeoff every agent carries. Start on Plus, keep prompts specific, and turn on Lockdown Mode for any account where a mistake would be expensive.