What is ChatGPT Work?
ChatGPT Work is the result of OpenAI merging three things that used to be separate: the ChatGPT desktop application, the Codex coding agent, and the plugin ecosystem. Launched on July 9, 2026, Work gives ChatGPT the ability to do work — not just talk about it.
Before Work, you used ChatGPT to get answers, Codex to write code, and plugins to connect external tools. Now, within the Work interface, you can ask it to draft a report, schedule it to run every Monday, have it pull data from your email via a plugin, edit the resulting document, and monitor the whole process from your phone. That is the pitch. For most use cases, it holds up.
The key shift is from single-turn responses to agentic loops. You give Work a goal, and it plans, executes, validates, and iterates until the goal is met or it needs your input. This is closer to having a junior employee than a search engine.
What can ChatGPT Work actually do?
Write code and manage projects
Codex is the coding layer inside Work. It runs in isolated cloud containers with your repository preloaded, can read and edit files, run terminal commands, execute tests, and create pull requests directly on GitHub. Multiple tasks execute simultaneously in separate containers — you can have Codex working on a bug fix and a new feature at the same time.
The /goal command takes this further. Type /goal Complete the migration from REST to GraphQL without stopping until all tests pass and Codex enters a persistent autonomous loop. It plans, acts, tests, reviews, and iterates — for minutes or hours — without needing you to steer each step. You can check progress, pause, resume, or clear the goal at any time. Most real-world /goal sessions for meaningful features run between 20 minutes and 2 hours.
Browse the web
Work includes a built-in browser in the desktop app. Instead of giving you a link to go read, ChatGPT can now navigate pages, read content, fill forms, and extract information directly. This is different from the old "search the web" feature — the agent actually controls a browser session and can take actions on what it finds.
Call 1,400+ plugins
As of the launch, over 1,400 plugins are available in the ChatGPT Work directory, covering email, calendars, CRM, spreadsheets, document management, and more. Plugins let ChatGPT connect to external services and take actions — sending emails, creating calendar events, updating a CRM record, pulling data from a database.
OpenAI migrated all plugins from the old app directory to the Work hub on July 9, 2026. You browse and install them directly from within ChatGPT.
Use reusable skills
Skills are reusable workflows that encode a multi-step process — similar to how you might teach a new hire a repeatable procedure. Skills follow the Agent Skills open standard (defined at agentskills.io), so they are portable across compatible agent platforms.
The difference between a plugin and a skill: a plugin connects ChatGPT to an external service (like Gmail). A skill teaches ChatGPT how to do a task (like "run a competitor analysis and format it as a one-pager"). You can find, install, and share skills in the same directory as plugins.
Schedule recurring tasks
Work includes a Scheduled Tasks feature. You can set up a task to run on a recurring schedule — daily, weekly, or at a specific time. For example: "Every Monday at 9 AM, pull last week's sales data from the CRM plugin, generate a summary report, and draft an email to the team."
This is where Work crosses from being a tool you use interactively into being a system that works in the background while you do other things. Some of the most successful AI automation projects we have documented follow exactly this pattern — set up a recurring task once, let it run, review the output.
Control your desktop from mobile
Codex Remote, which reached general availability on June 25, 2026, lets you monitor and control Codex tasks from the ChatGPT iOS or Android app. You can approve actions, change models, and view outputs from your phone. The tasks still execute in the cloud or on your desktop — the phone is a remote control, not the execution environment.
Talk instead of type
Voice mode is available in both Work and Codex on desktop (macOS and Windows), with iOS as a remote pairing device, as of July 23, 2026. You speak to ChatGPT, it responds in voice, and you can have a back-and-forth conversation about a task while it is being worked on. If you have never tried voice prompting, it changes how you interact with AI — natural conversation surfaces ideas and clarifications that typing does not.
How much does ChatGPT Work cost?
ChatGPT Work is included across all ChatGPT plans. The feature set varies by tier.
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | Codex (limited), ChatGPT web, no plugins |
| Go | $8/mo | Codex, higher limits than Free |
| Plus | $20/mo | Codex across web/CLI/IDE/iOS, GPT-5.5 models, plugins |
| Pro (5x) | $100/mo | 5x Plus rate limits |
| Pro (20x) | $200/mo | 20x Plus rate limits, highest individual usage |
| Business | $20–25/seat/mo | SAML SSO, MFA, workspace credits, min 2 users |
| Enterprise | Custom | SCIM, EKM, RBAC, audit logs, data residency |
Codex shares an "agentic usage limit" with ChatGPT for Excel and Workspace Agents. Plus and Pro users can buy extra credits at the cap. Since April 2, 2026, Codex is billed on API token usage (credits per million input/cached/output tokens) on Plus, Pro, Business, and Enterprise, replacing per-message pricing.
Typical real-world Codex cost runs about $100–200 per developer per month, varying by model, parallel instances, and whether fast mode is enabled.
Which GPT-5.6 model should you use?
ChatGPT Work runs on GPT-5.6, which has three tiers with different effort and speed tradeoffs:
| Model | Best for | Input per M tokens | Output per M tokens |
|---|---|---|---|
| Luna | Speed and cost-efficiency | $1 | $6 |
| Terra | Balanced tasks | $2.50 | $15 |
| Sol | Maximum reasoning power | $5 | $30 |
Sol is the default for Codex and appears in the model picker with configurable effort settings (Medium, High, Extra High). Higher effort means more reasoning tokens spent before an answer — better for complex problems, more expensive. For simple tasks, switching to Luna can cut costs by 80% with minimal quality loss.
Is ChatGPT Work safe for enterprise use?
Work includes enterprise governance controls. Admins can set permissions, restrict which plugins are available, enable audit logs, and require approval before the agent takes certain actions (like sending emails or calling external APIs). An auto-review feature runs before tool and API actions execute, adding a layer of automated safety checking.
For regulated industries, Enterprise tier adds data residency, encryption key management (EKM), and role-based access control. Codex sandboxed cloud environments run in isolated containers with configurable network access per workspace — so you can prevent agents from reaching the internet during execution if your security policy requires it.
That said, autonomous agents that take real-world actions are still new territory. The 95% enterprise AI pilot failure rate is driven partly by organizations skipping the governance and guardrails step.
ChatGPT Work vs alternatives: how does it compare?
ChatGPT Work is not the only agent platform, but it is the most integrated with the OpenAI ecosystem. Here is how it stacks up:
- vs. Claude with computer use: Claude can control a desktop, but lacks Codex's code execution sandbox, the
/goalpersistent loop, and the 1,400+ plugin directory. Claude is stronger for nuanced reasoning and long-context analysis; Work is stronger for coding and automation. - vs. Google Gemini's agentic features: Gemini has deeper integration with Google Workspace (Docs, Sheets, Drive), but its plugin ecosystem is smaller and it does not have a
/goal-equivalent for multi-hour autonomous runs. - vs. open-source agent frameworks (Hermes, etc.): Open-source gives you full control over data and execution, and you can run locally without per-token costs. But you give up the 1,400+ plugin directory, the cloud sandboxes, and the mobile remote control. For many builders, the answer is running both — ChatGPT Work for plugins and hosted convenience, an open-source agent for sensitive or cost-intensive work.
How to get started with ChatGPT Work
Pick a plan. Plus ($20/mo) is the sweet spot for individuals. Business ($20–25/seat) if you need SSO and team controls.
Install the desktop app. The desktop app (macOS or Windows) gives you the built-in browser, voice mode, and Codex. The web version is more limited.
Install plugins. Browse the plugin directory from within Work. Start with the services you already use — email, calendar, CRM. You do not need 1,400 plugins. You need 3–5 that match your actual workflow.
Try Codex on a real task. Give it a small bug fix or feature from your repository. Watch how it plans, executes, and creates a PR. This teaches you the pattern for larger tasks.
Set up one scheduled task. Pick something you do weekly — a report, a summary, a data pull. Schedule it in Work and let it run. This is where the real time savings come from.
Try
/goalon a scoped task. Start with something that should take 30 minutes — not an overnight project. You need to learn how/goalstops before you trust it with bigger work.
What this means for you
If you are a small business owner or operator, ChatGPT Work is the first product that puts the "agent" promise in a usable, package — it does things for you rather than just answering questions. The practical path is not to use everything at once. Pick one workflow that takes you 2–3 hours per week, set it up as a scheduled task with the relevant plugins, and let Work handle it. That is the test that tells you whether this matters for your business.
If you are a developer, Codex with /goal is the feature to try — it changes the unit of work from a single prompt to a full project. Start with a 30-minute scoped task, review the diff carefully, and scale from there. And if you want to run agents that work on your own infrastructure with full data control, you can pair ChatGPT Work with an open-source agent framework for the best of both worlds.
Related reading
FAQ
What is the difference between ChatGPT and ChatGPT Work? ChatGPT is a chatbot that answers questions. ChatGPT Work is an agent layer that includes the Codex coding agent, 1,400+ plugins, reusable skills, a built-in browser, scheduled tasks, voice mode, and mobile remote control. Work can take actions — not just respond to prompts.
Is ChatGPT Work free? ChatGPT Work features are included across all ChatGPT plans. Free users get limited Codex access. Plus ($20/mo) unlocks Codex across web/CLI/IDE/iOS, plugins, and GPT-5.5 models. Pro ($100 or $200/mo) offers 5x or 20x Plus rate limits. Business and Enterprise add team controls.
What is the /goal command in Codex?
/goal is a slash command that turns a Codex prompt into a persistent autonomous loop. You specify a high-level objective and a stopping condition, and Codex plans, executes, tests, and iterates until the condition is met — for minutes or hours — without needing your input at each step. You can pause, resume, or clear the goal at any time.
Can ChatGPT Work browse the internet? Yes. The desktop app includes a built-in browser that lets ChatGPT navigate pages, read content, fill forms, and extract information directly — not just search and link. This is available on macOS and Windows.
How many plugins does ChatGPT Work have? Over 1,400 plugins were available at launch on July 9, 2026, covering email, calendars, CRM, spreadsheets, document management, and more. All plugins from the old app directory were migrated to the Work hub.
Can I control ChatGPT Work from my phone? Yes. Codex Remote (GA June 25, 2026) lets you monitor and control Codex tasks from the ChatGPT iOS or Android app. You can approve actions, change models, and view outputs from your phone. Tasks execute in the cloud or on your desktop — the phone is a remote control.
Which GPT-5.6 model should I use? Sol (default for Codex) for maximum reasoning power. Terra for balanced tasks. Luna for speed and cost-efficiency — it costs 80% less than Sol. For simple tasks, Luna is usually enough. For complex coding or analysis, use Sol with High or Extra High effort.
Every claim here is traced to a primary source, dated, and listed under Sources. Research and drafting are AI-assisted; editing, verification and publication are human decisions, and a person is accountable for what appears on this page. How we work →
Discussion
0 comments