Contents
- 1. Bottom line — ChatGPT Work is a "work agent that builds finished deliverables"
- 2. What it can do — docs, sheets, slides, web apps
- 3. The unified app's three faces — Work / Codex / regular chat
- 4. Models and plans — what you can pick depends on your plan
- 5. Connectors — which apps and files it reaches
- 6. How to use ChatGPT Work — getting started and prompting tips
- 7. Honest caveats — what isn't settled yet
- Summary
- FAQ
ChatGPT Work is the "AI agent for work" that OpenAI announced on July 9, 2026, alongside its new GPT-5.6 model family (OpenAI announcement, Axios, TechCrunch). It isn't just a chat that answers questions — its defining trait is that it gathers context from your connected apps and files and builds "finished deliverables" like documents, spreadsheets, slides, and web apps. At launch its brain was GPT-5.6, but newer models have been added one after another since — GPT-6 Astra (September 3, 2026), then GPT-6 Sol and Luna (September 22, 2026) (OpenAI release notes). Think of Work not as a product tied to one model but as a workspace where the models you can pick keep changing, depending on your plan.
This article is built on OpenAI's own announcement, Help Center, and release notes, supplemented by launch-day reporting (Axios / TechCrunch / 9to5Mac / Thurrott). It covers what ChatGPT Work is, how it differs from regular ChatGPT and Codex, and which models each plan can use. Everything about models and plans was checked against OpenAI Help and the official docs (learn.chatgpt.com) as of September 25, 2026. We re-checked the same sources on September 26, 2026 and added how the Default setting in the model picker works. Where OpenAI hasn't documented something, we mark the confidence level explicitly.
An agent that returns "deliverables," not chat
1. Bottom line — ChatGPT Work is a "work agent that builds finished deliverables"
In Thurrott's words, ChatGPT Work is "an agent in ChatGPT that helps you take on more ambitious tasks," one that "gathers information across your apps and workflows to create finished materials like sheets, slides, docs, and web apps" (Thurrott). Three things stand out.
- It returns "deliverables," not "answers": regular ChatGPT replies in conversation; Work actually assembles real files (spreadsheets, slides, and so on).
- It executes multiple steps autonomously: Thurrott reports it can "break a complex project into smaller steps and keep working for hours." It sticks with long jobs rather than firing off a single reply.
- You choose the brain — and it gets swapped over time: at launch, Work ran on GPT-5.6 (OpenAI announcement / Thurrott). Since then GPT-6 Astra (September 3, 2026) and GPT-6 Sol and Luna (September 22, 2026) have been added, and you now pick a model for each job from the model picker in Work. Which ones you can pick depends on your plan. Section 4 has the details.
So it helps to think of ChatGPT Work as the "driver's seat tuned for office work" on top of "engines" like GPT-5.6 and GPT-6 Astra. The engine can be swapped; the driver's seat stays the same.
2. What it can do — docs, sheets, slides, web apps
The use case reporters consistently cite is everyday clerical work. TechCrunch describes a "workplace companion" that helps with "daily clerical tasks like drafting documents, spreadsheets, and presentations" (TechCrunch). Thurrott emphasizes that it builds finished materials from templates and reference files.
Aggregates data from connected sources and assembles it into a spreadsheet.
Creates presentations and documents from your reference files.
Can output simple web apps as finished deliverables too.
Breaks big jobs down and keeps at them for hours (Thurrott).
In short, Work goes beyond "look it up and stop" to "build it and hand it over." If GPT-Live, which talks with you in voice, is the "talking partner," Work is the "hands-on partner."
3. The unified app's three faces — Work / Codex / regular chat
The new ChatGPT desktop app released on July 9, 2026 brings Chat, Work, and Codex together in one app (OpenAI release notes). ChatGPT Work is positioned as one of those modes.
| Mode | Role | Who it fits | Where it runs |
|---|---|---|---|
| ChatGPT Work | The agent that builds deliverables — docs, sheets, slides, web apps | Business roles: ops, planning, sales | Web, phone, desktop app |
| Codex | Code and engineering work. Shows technical details | Engineers and developers | Desktop app, CLI, IDE. Cloud is also available on web and phone, separately from remote access to a PC |
| Regular chat (Chat) | Fast conversation and everyday questions | Research, drafts, brainstorming | Web, phone, desktop app |
9to5Mac reports that inside the desktop app, "ChatGPT Work abstracts away (hides) the technical details that Codex shows" (9to5Mac). In other words, Work is the "non-engineer version" of Codex — a doorway to the same underlying execution power, aimed at business deliverables without making you think about code. To grasp how Codex and Work relate, it helps to understand where Codex sits.
The difference from regular chat (Chat) isn't only what it returns. According to OpenAI Help, three things differ: (1) usage — Work shares its allowance with Codex and is counted separately from Chat's message limits; (2) models — GPT-6 Sol and Luna, and GPT-5.6 Terra and Luna, can't be selected in regular Chat, only in Work and Codex (and on Plus, Astra is also limited to Work and Codex); (3) task length — Chat is for fast conversation, Work for longer multi-step jobs and finished deliverables (OpenAI Help).
Can the same Work conversation continue after you close your PC?
Coordination in the cloud
The official "Local computer access with Work Cloud" feature lets you continue the same conversation on web, phone, and desktop while using files and tools on an authorized PC. Availability depends on the workspace and rollout.
No local work without the PC
Steps that use the PC require it to be online and connected, with the desktop app running and signed in. If the PC is unavailable at the start of a new turn, eligible tasks can continue in the cloud, but cannot use that PC's files or tools.
This applies to new tasks created after the feature is enabled. The official Enterprise guide says to start with Cloud selected in the desktop app. Existing tasks retain their original mode, and a running turn cannot automatically switch from local execution to Cloud. Syncing a conversation does not upload the entire local folder. This guide is not a guarantee that the feature is available to every personal account.
On October 8, 2026, we read the full official Local computer access for Work Cloud and dots, Remote connections, and Projects and chats documentation. For differences from Codex Remote and Cloud, see our comparison of PC sleep and execution locations.
4. Models and plans — what you can pick depends on your plan
Work's model lineup grows and changes every time OpenAI ships a new model. So rather than memorizing names, it helps to think in three tiers of roles, which won't go stale: (1) the top tier (flagship) — for hard research, analysis, and complex problems, and it drains the allowance fastest; (2) the balanced tier — capability and efficiency together for complex work; (3) the low-cost tier — focused, high-volume jobs like summarizing, extracting, and classifying. When a new model arrives, first check which tier it belongs to, and the way you choose stays the same.
As of September 25, 2026, tier (1) is GPT-6 Astra, tier (2) is GPT-6 Sol and GPT-5.6 Sol, and tier (3) is GPT-6 Luna and GPT-5.6 Luna, with GPT-5.6 Terra ("a balance of speed, capability, and cost for everyday work") in between. In OpenAI's words, GPT-6 Sol is for "complex coding and agentic workflows" and GPT-6 Luna is its "most efficient model for focused, high-volume tasks." GPT-6 Sol and Luna can be picked only in Work and Codex, not in regular Chat, and the three GPT-5.6 models remain available during the rollout (official "Models" docs / OpenAI Help). Which ones you get, and how much, depends on your plan. For the current models across vendors, see our list of current models and knowledge cutoffs; for GPT-6 Sol and Luna in depth, see the GPT-6 Sol and Luna release explainer.
Models available in Work by plan (OpenAI Help and official docs, as of September 25, 2026)
API reference prices (per 1M tokens, input/output, from the OpenAI pricing page as of September 25, 2026): GPT-6 Astra $10/$50, GPT-6 Sol $2/$10, GPT-6 Luna $0.10/$0.50, GPT-5.6 Sol $5/$30, GPT-5.6 Terra $2/$12, GPT-5.6 Luna $0.20/$1.20 (Terra and Luna revised July 30, 2026; GPT-5.6 Sol is $4/$20 on promotional pricing for three months from August 21, 2026). Note: Work inside the ChatGPT app runs on your plan's included allowance (shared with Codex), separate from API prices. The Sol price cut also applies to Work usage paid in credits, while the subscription allowance itself is unchanged (OpenAI Developer Community).
Astra drains the allowance faster. Work and Codex share one allowance, with five-hour and weekly limits depending on the plan. OpenAI Help's estimates (messages per five hours for local tasks; actual usage varies widely by task) put Plus at 5–45 for GPT-6 Astra, 10–100 for GPT-5.6 Sol, 15–150 for GPT-6 Sol, 25–200 for GPT-5.6 Terra, and 350–3,000 for GPT-6 Luna (official pricing page, as of September 25, 2026). In that order, each message draws down less of the allowance. At the same time, OpenAI notes that "Astra at Low effort can outperform Sol at High effort" (meaning GPT-5.6 Sol) and suggests trying Astra at Low or Medium for work you were happy with on Sol at High (OpenAI Help). In other words, "a higher-tier model at a low setting" versus "a mid-tier model at a high setting" is worth comparing on your real work.
What changed when (from OpenAI announcements and release notes)
| Date | Change |
|---|---|
| July 9, 2026 | ChatGPT Work announced (powered by GPT-5.6). On web and phone, rolling out first to Pro, Enterprise, and Edu, with Plus and Business within days. The new desktop app (Mac, Windows) launched globally the same day |
| July 16, 2026 | Work chats run in the cloud started syncing across web, phone, and desktop |
| August 21, 2026 | The three-month GPT-5.6 Sol price cut also applied to Work usage paid in credits |
| September 3, 2026 | GPT-6 Astra announced, starting with a limited set of organizations |
| September 4, 2026 | OpenAI on X: Astra available in Work and Codex for Pro, Enterprise, and Business Premium; Plus and Business could take a few days |
| September 22, 2026 | GPT-6 Sol and Luna arrive in Work and Codex (not selectable in regular Chat) |
| September 23, 2026 | On web and phone, you can now talk to Work by voice (Live) and hand it tasks |
| As of September 25, 2026 | Plus can use Astra in Work and Codex (limited usage). GPT-6 Pro (powered by Astra) in regular Chat is for Pro, Business, and Enterprise |
| October 14, 2026 (scheduled) | GPT-5.5 retires from ChatGPT, Work, and Codex (the API is not affected). The replacement is GPT-6 Sol (GPT-6 Luna on Free and Go) |
Sources: OpenAI release notes, ChatGPT Work announcement, GPT-6 Astra announcement, OpenAI Developer Community post, OpenAI Help, official "Models" docs
🟡 OpenAI's wording on the Free and Go plans is inconsistent. The launch blog post said Work on web and phone starts with paid plans, but also said that in the new desktop app "Chat, Work, and Codex are available to users on all plans, including Free." As of September 25, 2026, however, OpenAI Help says Work on web and phone is for "eligible paid plans" and in the desktop app is available "when included for your plan and workspace," and its table of models available in Work lists only Plus, Pro, Business, and Enterprise. For Free and Go, OpenAI Help's table says GPT-5.6 Terra in Codex while the Codex pricing page says GPT-6 Luna in the desktop app (rolling out) — neither states what Work itself offers. Check your own app to see whether Work is available on a free plan.
5. Connectors — which apps and files it reaches
Work's value lies in "being able to pull in your context." ChatGPT can reach files in external services through connectors (app integrations), supporting Google Drive, SharePoint, Dropbox, Box, OneDrive, Gmail, Slack, and GitHub, among others (OpenAI Help). For enterprises, custom integrations are also possible via MCP (Model Context Protocol). Alongside Work's launch, the App Directory was replaced by the Plugin Directory as the place to add integrations (existing connections keep working; OpenAI release notes).
🟡 Confidence note
The connector list above is for ChatGPT's integration platform overall, not a list confirmed as "ChatGPT Work only." Exactly which connectors Work can use, and how far, should be checked against updated official docs (this table is compiled from reporting plus the existing connector spec).
On privacy, connector data for Business, Enterprise, and Edu is not used for model training by default (OpenAI Help). Since this is an agent meant to handle work files, that "not used for training by default" default matters in practice.
6. How to use ChatGPT Work — getting started and prompting tips
First, some context: ChatGPT Work pays off in work that is more "build" than "look up" — ops / back office (mass-producing routine reports, aggregations, and meeting notes), planning / sales (first-draft slides and proposals from reference materials), and small internal tools (a simple web app without writing code). Here's the actual workflow, in the order (1) getting started → (2) how to prompt → (3) tips.
Step 1: Getting started (4 steps)
On web or phone, pick Work at the top of the screen. In the desktop app, choose ChatGPT at the top left, then switch to Work at the top.
In Settings → Plugins, authorize Google Drive, Slack, etc. on the consent screen (a few minutes).
Be specific about "what, in which format, from which data."
Check the output and refine with follow-ups. Always verify numbers and citations.
Step 2: How to prompt (examples)
Work gets more accurate the clearer you are about "what, in which format, from which data." Example instructions by use case:
| Deliverable | What the instruction looks like |
|---|---|
| Spreadsheet | "Aggregate the sales data in the connected Google Drive by month, and make a spreadsheet with a month-over-month column and a totals row." |
| Slides | "Based on this document, make an executive proposal deck in 10 slides or fewer. Put the conclusion on slide 1 and next steps at the end." |
| Web app | "Build a simple internal-form web app that takes an inquiry as input and lists the submissions." |
| Document | "Turn these meeting notes into a one-page report: a heading, three key points, and an action-item list." |
Step 3: Tips for good results
- Pick the model for the task: the top tier (GPT-6 Astra as of September 2026) for hard research and analysis, the balanced tier (Sol) for everyday deliverables, and the low-cost tier (Luna) for routine extraction, classification, and summaries. The top tier drains the allowance faster, so follow OpenAI's advice and start at Low or Medium (OpenAI Help). When new models arrive, just slot them into the same three tiers. If you are unsure, pick Default in the model picker: it puts models and reasoning levels on one slider, and the official docs illustrate six presets (Luna High, Sol Light, Sol Medium, Astra Light, Astra Medium, Astra Extra High) with Sol Light as the starting point (some paid plans omit Astra Extra High; as of September 26, 2026, official docs “Models”).
- Give reference files: connecting or attaching a template or past materials sharply improves formatting and consistency.
- Specify the format: state the slide count, columns, tone, and output type (spreadsheet, slides, etc.).
- Ask in stages: for big jobs, split into "outline first → flesh it out → polish" for stability.
- 🔴 Always review the output: because it builds autonomously, numbers, citations, and sensitive data need a human check — don't submit as-is.
- Start with least privilege: connect only the apps and permissions you actually need.
Note: screen and step details may change by version. For the latest steps, check the official Help.
7. Honest caveats — what isn't settled yet
Two and a half months after launch, the model lineup in Work has already grown twice (GPT-6 Astra, then GPT-6 Sol and Luna), and GPT-5.6's predecessor GPT-5.5 is scheduled to retire from ChatGPT, Work, and Codex on October 14, 2026. Read this assuming model names will keep changing; here is the confidence as of September 25, 2026, split by level.
- 🟢 Confirmed by OpenAI: announced July 9, 2026 alongside GPT-5.6 / an agent that builds deliverables (docs, sheets, slides, reports, Sites) / built on Codex technology / runs on web, phone, and the desktop app (OpenAI announcement). Plus, Pro, Business, and Enterprise can pick from several models in Work, including Astra (OpenAI Help).
- 🟡 OpenAI's wording is inconsistent: whether Free and Go can use Work (the launch blog and the Help table don't match). Astra usage is published only as wide estimates.
- 🟡 Unsettled: Work's exact connector (plugin) coverage, and the edit/export formats for deliverables.
- 🔴 Don't over-trust: since the agent builds deliverables autonomously, numbers, citations, and sensitive data still need human review. Especially when pulling in external data, don't skip verifying outputs and designing permissions carefully.
Summary
- ChatGPT Work is a "work agent that builds deliverables": it assembles docs, sheets, slides, and web apps from the context of your connected apps and files (announced July 9, 2026).
- One of the unified app's three modes: Work (deliverables) / Codex (technical, desktop, Cloud, and other entry points) / regular chat (conversation). Work is the "business version" that hides Codex's technical detail (9to5Mac). Its allowance and model choices differ from Chat.
- Choose models by role: think in three tiers — top, balanced, and low-cost. As of September 25, 2026, Work offers GPT-6 Astra, Sol, and Luna plus GPT-5.6 Sol, Terra, and Luna. Astra is limited on Plus and Business Standard, while Pro and Business Premium can spend their full allowance on it (OpenAI Help). OpenAI's wording on Free and Go is inconsistent.
- Its strength is "your context": connectors (plugins) like Drive/SharePoint/Slack plus MCP; enterprise data isn't used for training by default.
- Human review is the premise: because it builds autonomously, verification and least-privilege are the practical keys. Confirm details against official updates.
In one line — ChatGPT Work is a step from "AI that answers" to "AI that works." It started on GPT-5.6, and its engine has since been swapped for GPT-6 Astra and GPT-6 Sol and Luna. The driver's seat stays the same while the engine gets newer — worth knowing as the product that wires those models directly into producing office deliverables.
FAQ
Q1. How is ChatGPT Work different from regular ChatGPT?
Regular ChatGPT returns "answers" in conversation, while Work is an agent that builds "deliverables" like docs, sheets, slides, and web apps. It gathers context from connected apps and files and works autonomously across multiple steps for hours (OpenAI announcement). On top of that, its allowance is separate from Chat (shared with Codex), and GPT-6 Sol and Luna, and GPT-5.6 Terra and Luna, can't be selected in regular Chat — you pick them in Work. On Plus, GPT-6 Astra is also available only in Work and Codex (OpenAI Help).
Q2. How is it different from Codex?
Both use the same model family (OpenAI's GPT-6 and GPT-5.6 models) and a shared allowance, but Codex is for code and engineering work and shows technical details. Alongside the desktop app, CLI, and IDE, Cloud tasks can be started and continued on web and phone. Work hides that technical detail, specializes in business deliverables, and runs on web, phone, and desktop (official Codex Cloud documentation / OpenAI Help). Remotely controlling a PC from your phone is different from execution in Cloud. Engineers start with Codex; ops and planning roles start with Work.
Q3. Can I use it on free? Which model?
OpenAI's wording is inconsistent. The launch blog post said every plan, including Free, could use Work in the new desktop app, but as of September 25, 2026, OpenAI Help's table of models available in Work lists only Plus, Pro, Business, and Enterprise. For Free and Go, OpenAI only says which models Codex offers (OpenAI Help says GPT-5.6 Terra; the pricing page says GPT-6 Luna in the desktop app). Paid plans can choose from several models, including the top-tier GPT-6 Astra.
Q4. Which apps can it connect to?
ChatGPT's connectors support Google Drive, SharePoint, Dropbox, Box, OneDrive, Gmail, Slack, GitHub and more, and enterprises can add custom integrations via MCP (OpenAI Help). Work's exact coverage isn't individually confirmed, so checking updated official docs is safest (🟡).
Q5. Is it safe to feed it work files?
For Business, Enterprise, and Edu, connector data is not used for model training by default (OpenAI Help). Even so, start with the minimum apps and permissions, and handle sensitive data per your organization's policy.
Q6. When can I use it?
Announced July 9, 2026, it rolled out on web and phone first to Pro, Enterprise, and Edu, then to Plus and Business within days, and the new desktop app (Mac, Windows) launched the same day (OpenAI announcement). GPT-6 Astra was announced on September 3, reached Work for Pro, Enterprise, and Business Premium the next day, and later expanded to Plus. On September 22, GPT-6 Sol and Luna joined Work as well. Availability may differ by region, plan, and workspace settings.
Q7. How do I use GPT-6 Astra on Plus?
Pick it as the model in Work or Codex. Plus can use Astra in Work and Codex, but GPT-6 Pro (powered by Astra) in regular Chat is for Pro, Business, and Enterprise and isn't included in Plus. Astra usage on Plus is limited, and you can add more with credits. If it doesn't show up in the desktop app, update to the latest version via "Check for Updates" in the menu (OpenAI Help).
Related articles
- What is ChatGPT Sites — the feature that publishes sites you build in Work; what chatgpt.site really is, pricing, and caveats before using it in production
- GPT-5.6 release explainer — details of the Sol/Terra/Luna models you can pick in Work
- GPT-6 Astra release explainer — pricing and availability of OpenAI's flagship model that joined Work
- GPT-6 Sol and Luna release explainer — the balanced and low-cost models that joined Work and Codex
- What is GPT-Live — the same-day full-duplex voice; the "talking partner" contrast
- Claude Code vs Codex — where Codex, Work's engineering counterpart, sits
- AI's Impact on the Consulting Industry — what deliverable-building AI did to an industry that lived on making decks
- How Does AI Widen the Ability Gap Among Office Workers? — how the gap opens between colleagues who master such tools and those who don't
For editing an open Word document, see our ChatGPT for Word guide. It covers installation, proofreading and summary prompts, and pricing; the add-in has separate chat history and does not inherit ChatGPT memory.
