OpenAI offers several agent products with computer access, tools, and overlapping models. Dots, announced at DevDay 2026 on September 29, 2026, are always-on personal agents powered by GPT-6 Astra. ChatGPT agent, launched July 17, 2025, is an agent mode you turn on inside a chat. Codex is the company's software-engineering agent. ChatGPT Work is the sibling agent for longer knowledge work. Workspace agents are shared team agents on Business, Enterprise, and Edu.

This comparison uses OpenAI’s published descriptions. OpenAI marks the 2025 ChatGPT agent and Codex launch posts as outdated; their launch details below are historical, not current availability or quota guidance.

#What is each product, in OpenAI's words?

Dots. "Always-on agents in ChatGPT that can take on ongoing work and keep making progress between conversations." Each dot has its own cloud computer, works across connected apps, and is powered by GPT-6 Astra. You give it a goal and define what it can do on its own. It brings results back for review and turns to you for decisions that need judgment.

ChatGPT agent (2025). "ChatGPT now thinks and acts, proactively choosing from a toolbox of agentic skills to complete tasks for you using its own computer." OpenAI described it as a unified system that combined Operator's website interaction, deep research's synthesis, and ChatGPT's conversation. Users activated it by selecting agent mode from the tools dropdown in any conversation.

Codex. OpenAI's original Codex launch (May 16, 2025) called it "a cloud-based software engineering agent" that writes features, fixes bugs, and proposes pull requests in an isolated environment preloaded with a repository. The current help article still says "Codex remains dedicated to software development and technical work."

ChatGPT Work. "An agent designed for longer, multi-step work and finished deliverables" such as documents, spreadsheets, presentations, reports, and Sites. Work follows the same usage structure as Codex.

Workspace agents. "Codex-powered agents for teams," introduced in research preview on April 22, 2026. The page now carries a general-availability notice for Business, Enterprise, and Edu. OpenAI called them an evolution of GPTs: shared agents that run in the cloud and can be used together in ChatGPT or Slack.

Dots can start or manage tasks in Codex and ChatGPT Work. That is the join OpenAI has documented. It has not said a personal Dot replaces workspace agents.

#How they differ on the points OpenAI has documented

Topic

Dots (Sep 2026)

ChatGPT agent (Jul 2025)

Codex / Work

How you start it

Create a named primary dot on desktop ChatGPT

Select agent mode in a conversation

Switch to Work or Codex, or start a cloud task

Persistence

Keeps making progress between conversations; proactive research in the background

Runs the task you started; you can schedule completed tasks to recur

Cloud tasks can continue while a laptop is closed

Model (as published)

GPT-6 Astra

Unified agentic system at launch

GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna (plan-dependent)

Computer

Own cloud computer; optional local computer via desktop app

Own virtual computer with visual browser, text browser, terminal

Cloud environments, local folders, desktop apps (permissions apply)

Apps

Plugins, "over 4,000 apps"; Slack and Teams messaging

Connectors such as Gmail and GitHub; browser takeover for logins

Plugins, MCP, GitHub, local tools

Channels

ChatGPT web, mobile, desktop; Slack; Teams; voice call; Help Center describes US Pro texting beta, while launch post says coming soon

Inside ChatGPT; phone notification when a task finishes

Desktop app, web, mobile (Codex Cloud); CLI

Usage accounting

Dot conversations do not count toward ChatGPT limits; Work/Codex tasks do

Launch: 400 messages per month for Pro, 40 for other paid users

Shared Work and Codex pool; credits after included usage

Who gets it at launch

Pro, Business Premium, Enterprise (admin on); 18+; Pro excludes EEA, Switzerland, UK

Pro, Plus, and Team first; Enterprise and Education later; EEA and Switzerland still being enabled at launch

Paid plans; workspace controls for Enterprise/Edu

Figures in that table are as published on the cited OpenAI pages. ChatGPT agent's 400 / 40 monthly message caps are from the July 17, 2025 launch post. OpenAI has not republished those exact caps next to Dots.

#Persistence and initiative

The sharpest sourced split is when work continues.

Dots are built to hold ongoing responsibility. Help-center copy says a dot "can take on ongoing responsibility and keep making progress between conversations." When you are not talking to it, OpenAI says it looks for ways to help with proactive research: read-only tools on connected sources, private notes, no sending messages, no changing app content, no controlling a browser or computer until a follow-up passes the usual action checks.

ChatGPT agent was framed as a task you start. You describe a job ("plan and buy ingredients," "analyze three competitors and create a slide deck"). You can interrupt, take over the browser, or stop. You could schedule a completed task to recur, such as a weekly metrics report. OpenAI did not describe that 2025 product as a named companion that keeps a life-long context and pings you in Slack.

Codex and Work sit in the middle on persistence. Codex Cloud and cloud Work tasks continue while a computer is asleep. Event-triggered Work tasks can fire on new Gmail, Slack, or GitHub activity. Those are still tasks with a schedule or a trigger, not a single always-on identity.

#Identity, naming, and channels

A Dot is a person-shaped product. You name it during setup. The default handle is @yourname-dot; a custom name becomes @yourname-agentname. You pick a character or a pet. After setup you message it in ChatGPT, including mobile once access is available, and on Slack or Teams, with Teams limited to an invite-only alpha in the Enterprise admin guide. OpenAI says you can hop on a voice call. The Help Center describes a limited US Pro texting beta using a third-party provider; the launch post still says texting is coming soon. A dot cannot initiate calls to you at launch, and it cannot have its own standalone email address at launch.

ChatGPT agent had no separate name, avatar, or Slack identity in the launch post. You stayed inside the ChatGPT composer.

Codex is a mode and a set of surfaces (CLI, desktop, cloud, IDE), not a character. Workspace agents are named shared agents for a team, closer to Dots on identity, with a different buyer: a workspace, not a personal Pro seat.

#Coding versus general work

OpenAI's developer example for Dots is explicit: a dot watches customer feedback, scopes bugfixes, builds and tests them, and brings complete PRs with videos. Help says a dot can create cloud tasks in Codex environments if you create the environment in Codex first. Local computer access, when you allow it, lets a dot create Work or Codex tasks and use local skills.

Codex remains the product OpenAI tells you to open for "write or debug code, run tests and commands, review changes, or work with a repository." Work is for research and finished documents. ChatGPT agent could run code in a terminal and produce slideshows and spreadsheets, but it was not a git-native coding agent with a CLI and pull-request review pane.

So the sourced split is: Codex is the coding surface; a Dot can delegate to it. The sources checked here do not provide a Dots-versus-Codex coding benchmark.

#Safety posture OpenAI has described

OpenAI documents action checks and approval requirements for these products. The details differ, and some authorizations can be given in advance.

ChatGPT agent (2025) required explicit confirmation before purchases, Watch Mode (active oversight) for sending emails, and refusals for high-risk tasks such as bank transfers. Users could take over the browser. OpenAI treated that launch as High capability in biological and chemical risk under its Preparedness Framework.

Dots add Custom Rules, Auto-review (a separate check before actions such as sending email or changing files), and safety monitoring that can pause work. Changing a password or transferring money is handed back to you. Purchases with a card saved on a merchant site require approval. Privacy and admin controls are documented in a dedicated Dots safety post.

Codex and Work inherit workspace browser-use, computer-use, and plugin admin controls. Those are Enterprise settings, not Dot-specific rules.

#Plans and rollout, as of September 29, 2026

Dots: Pro (excluding the EEA, Switzerland, and the UK at launch), Business Premium in all supported ChatGPT regions, Enterprise beta off by default. Create on desktop or desktop web only. Age 18 or older. First dot included on Pro or Business Premium. See how to get Dots.

ChatGPT agent at launch: Pro, Plus, and Team the first week; Enterprise and Education in the following weeks; EEA and Switzerland still being enabled. Operator.chatgpt.com was scheduled to sunset.

Codex / Work: paid ChatGPT plans, with Enterprise and Edu workspace toggles for Work Cloud, Work Local, and Codex Local.

Workspace agents: the current GA notice names Business, Enterprise, and Edu. The April preview also included Teachers and scheduled credit-based pricing to start May 6, 2026.

#What OpenAI has not compared

The sources checked here do not provide a head-to-head latency or quality evaluation, establish the current placement of agent mode next to a Dot, or give extra-Dot prices. Treat any ranking of "which is smarter" as unsourced unless it comes from a later first-party eval.

#What to watch next

Whether agent mode is retired into Dots, whether a Dot can be provisioned as a workspace agent, and how Dots usage is billed after the first month will decide how these products sit in one ChatGPT sidebar. Until OpenAI writes that down, the honest map is the one above: always-on companion, in-chat agent mode, coding agent, knowledge-work agent, shared team agent.