Key takeaways
- ChatGPT Work is a multi-step agent for knowledge work: Slack, Teams, Google Drive, SharePoint, local files, and a built-in browser—delivering spreadsheets, docs, slides, and web apps.
- Desktop three modes: Chat (short Q&A), Work (knowledge delivery), Codex (coding)—available on Mac/Windows for all plans; web/mobile Work is rolling from Pro/Enterprise to Plus/Business.
- ChatGPT Work pricing: usage follows the same Codex credit system; Plus ~$20/mo, Business ~$20/user/mo (annual), Enterprise custom (often ~$60/user/mo from 150 seats).
- Codex tutorial:
npm i -g @openai/codexorbrew install --cask codex; pick Codex in the desktop app or runcodexin a terminal signed into ChatGPT. - Model IQ is not the divider—workflow entry is: Work owns deliverables, Codex owns the repo; long agents belong on an isolated Cloud Mac running 24/7.
Conclusion: Work owns delivery, Codex owns code—do not force one mode to do both
Model capability is not the divider—workflow entry is.
On July 9, 2026 OpenAI launched ChatGPT Work alongside GPT-5.6. Work is positioned as a partner that finishes familiar multi-step tasks—budget analysis, competitor journey mapping, research through marketing assets. It shows a plan for review before acting, asks approval for sensitive steps, and can keep projects moving on a schedule while you are offline.
Codex remains the coding entry: terminal CLI, IDE plugins, desktop Codex tab, and cloud chatgpt.com/codex. Both Work and Codex consume more quota than everyday chat under the same credit structure. First rule of this guide: open Codex for code, Work for knowledge work, Chat for quick Q&A—do not write iOS projects in Work or expect Codex to format a board deck.
1. Why a separate Work mode in 2026?
The old problem was not a dumb model but broken context and execution chains. Typical failures:
- Asking Chat for a “competitor PPT” and getting only an outline while you copy Slack and Drive data by hand.
- A long task dying mid-run when the laptop sleeps or the topic switches and all business context is lost.
- Teams running Work on personal Plus accounts with no group spend controls at month end.
- Throwing coding into Work when GitHub PRs and local builds need Codex instead.
ChatGPT Work tries to fix this with connectors for Slack / Microsoft 365 / Google Drive, desktop access to local files and a built-in browser, and Enterprise spend controls in the Admin Console. Virgin Atlantic reported structured competitor journey analysis dropping from weeks to hours—the value is end-to-end orchestration, not a longer chat thread.
2. Chat / Work / Codex: classify before you compare
2.1 Chat (conversation mode)
Short Q&A, brainstorming, quick polish. Message rate is plan-limited; great for second-level interaction, poor for “research 20 sites then ship a shareable Excel.”
2.2 Work (knowledge-work mode)
Per the product page, core capabilities include:
- Cross-tool context: Slack, Teams, Google Drive, SharePoint, plus custom Team Agent plugins in Enterprise.
- Deliverables: spreadsheets, documents, slides, interactive web apps that stay live as source data updates.
- Plan first: complex jobs show steps; important actions wait for your approval.
- Scheduled runs: recurring workflows for weekly reports and monitoring.
- GPT-5.6 tiers: Sol for fuzzy goals, Terra for daily professional work, Luna for high-volume low cost (see our GPT-5.6 coding model guide).
2.3 Codex (coding mode)
Four surfaces, one brand (see the openai/codex repo):
- Codex CLI: terminal agent via
npm i -g @openai/codex, ideal over SSH on a server or Cloud Mac. - Desktop Codex tab: same app as Work; Windows adds Computer Use (region-limited).
- IDE plugins: VS Code, Cursor, JetBrains.
- Codex Web: cloud agent at
chatgpt.com/codex.
3. ChatGPT Work pricing and quotas (July 2026)
Key points from the business pricing page and Work announcement (USD; verify with OpenAI):
| Plan | Price | Work / Codex availability | Quota notes |
|---|---|---|---|
| Free | $0 | Desktop Chat / Work / Codex; web Work limited | Low ceiling; complex Work hits limits fast |
| Plus | ~$20/mo | Web/mobile Work rolling out | Includes Codex; long Work tasks burn faster |
| Pro | Higher personal tier | Web Work shipped first | Higher included usage |
| Business | ~$20/user/mo annual; ~$25 monthly | 2+ seats; SSO and spend analytics | Extra credits available; no training on business data by default |
| Enterprise / Edu | Contact sales; market ~$60/user/mo+ | Full features + SCIM/EKM/residency | Group and personal spend controls; credit add-ons with approval |
Billing logic: OpenAI states Work uses more than typical chat and bills under the same credit system as Codex—more complex tasks, more tool calls, and higher GPT-5.6 Sol reasoning burn credits faster. Enterprise admins can set per-person overrides for high-impact roles. Heavy multi-hour Work on Plus feels like “subscription is not enough” because you are buying agent minutes, not message counts.
4. Core comparison: Work vs Codex vs Claude Code vs Cursor
| Tool / approach | Entry | Execution | Context | Cost | Permission boundary | Best fit |
|---|---|---|---|---|---|---|
| ChatGPT Work | Desktop/web Work mode | Cross-SaaS + local files; Office-style deliverables; not Git-native | Company Drive/Slack + team plugins | Plan credits; long jobs cost more | Enterprise admin controls connectors and spend | Product, ops, strategy, consulting |
| Codex CLI / desktop Codex | Terminal / IDE / desktop tab | Edit code, run shell, open PRs; Computer Use on Windows desktop | Repository + terminal environment | Shares Codex quota with Work | Sandbox / user-approved commands | Engineers, DevOps, full-stack |
| Claude Code | Terminal CLI | Long-session refactors, Hooks, deep MCP integration | 200K + MCP expansion | Anthropic subscription / API | Self-managed Hooks and tools | Terminal-first teams valuing Skills |
| Cursor IDE | VS Code–family IDE | Inline completion + multi-file agent edits | Open files + indexed repo | Cursor subscription + model API | IDE sandbox | Daily pair programming, frontend iteration |
Running Work research and Codex patches on a Mac that sleeps when closed is exactly why Cloud Mac dual-agent isolation matters: Work outputs live in Drive while Codex edits an isolated worktree.
5. Codex tutorial: install through first successful run
5.1 Install Codex CLI (macOS / Linux / Windows)
⚠️ Package name must be @openai/codex—the unscoped codex npm package is an unrelated legacy project.
# Method A: npm (requires Node.js 22+)
npm install -g @openai/codex
# Method B: Homebrew (macOS)
brew install --cask codex
# Method C: official script (Mac/Linux)
curl -fsSL https://chatgpt.com/codex/install.sh | sh
# Method D: GitHub Release binary for your platform
# https://github.com/openai/codex/releases
5.2 Sign in and authenticate
# After launch choose Sign in with ChatGPT (uses Plus/Pro/Business quota)
codex
# Or API key (extra setup — see developers.openai.com/codex/auth)
export OPENAI_API_KEY=sk-...
Plan entitlements are documented in the Codex in ChatGPT help center.
5.3 Desktop app: switch Chat / Work / Codex
- Install the unified Mac or Windows client from the OpenAI download page.
- Sign in with ChatGPT (Enterprise via SSO).
- Pick Work for knowledge tasks or Codex for coding; use Chat for quick Q&A.
- Settings → model default: Terra daily, Sol for hard problems, Luna for batch preprocessing.
- Settings → connectors: authorize Google Drive, Slack, GitHub as needed (Work vs Codex differ).
5.4 First Codex task example
# At repository root
cd ~/projects/my-app
codex
# Example prompt inside the terminal session:
# "Read README, install deps, run tests, fix failures, do not change public API."
For long jobs SSH into a Cloud Mac and run codex while the desktop Work tab handles requirements docs. See our Claude Code Skills ranking for terminal-stack comparisons.
6. Scenario decision matrix
| Your scenario | First choice | Alternative | Avoid |
|---|---|---|---|
| Competitor analysis → dataset → executive deck | ChatGPT Work + Sol | Chat + manual export | Codex CLI |
| Fix bugs, open PRs, run CI | Codex CLI / desktop Codex | Claude Code | Work mode |
| Write and refactor React inside the IDE | Cursor + GPT-5.6 Terra | Desktop Codex | Work |
| Weekly report automation across Slack + Drive | Work + scheduled task | Custom GPT | Chat only |
| Team-wide spend and connector governance | Business / Enterprise | Business + credit add-ons | Each employee on personal Plus running Work |
| Real iOS / Xcode device builds | Codex on Cloud Mac | Claude Code on Cloud Mac | Windows laptop without macOS node |
7. Recommended stacks
Stack A — product / strategy (non-engineering)
Chat: Terra for daily Q&A
Work: Sol for competitor / financial models + Drive/Slack connectors
Codex: not used
Plan: Business (2+ seats) + admin spend caps
Stack B — individual engineer
Chat: Terra drafts
Work: Sol for RFCs and technical design docs
Codex CLI: Terra/Sol on Cloud Mac for code + tests
Plan: Plus or company Business
Stack C — engineering team
Work: Sol drafts release notes and changelog
Codex: GitHub connector + Cloud Mac worktree isolation
Claude Code: long refactors (separate directory from Codex)
Cursor: Terra for IDE daily work
Plan: Enterprise + group Codex/Work credits
Common pitfalls
- Pitfall 1: Work equals unlimited chat. Long agents bill credits; Sol plus many connectors can burn days of Plus quota in one run.
- Pitfall 2: writing production code in Work. No first-class Git support—switch to Codex or Claude Code.
- Pitfall 3: wrong npm package.
npm i -g codexis not OpenAI Codex; use@openai/codex. - Pitfall 4: ignoring Node 22. Confirm
node -v≥ 22 before npm install, or use Homebrew / binaries. - Pitfall 5: 24/7 Codex on a laptop. Sleep and network changes kill agents faster than model choice; move to Cloud Mac first.
Seven rollout steps
- Inventory task types: classify 80% of team work as Chat, Work, or Codex.
- Pick a plan: Plus to trial Work; Business for 2+ with SSO and spend panel; Enterprise for 150+ seats.
- Install the unified desktop app on Mac/Windows and confirm all three modes appear.
- Install Codex CLI per §5: sign in with
codexon Cloud Mac or dev machine and run a README + test smoke task. - Configure connectors: Work → Drive/Slack; Codex → GitHub; Enterprise uses admin allowlists.
- Set model defaults: Terra daily, Sol on a weekly hard-task budget, never default Sol for one-line email.
- Two-week review: log credit burn, success rate, and human interventions—tune the stack with data, not gut feel.
FAQ
What is the difference between ChatGPT Work and Chat?
Chat is for short interaction; Work runs multi-step projects with business tools and local files, producing shareable deliverables under the same credit system as Codex. One complex Work job can equal dozens of normal chats.
Can free users use ChatGPT Work?
Yes on desktop Mac/Windows—Chat, Work, and Codex are open to all plans including Free. Web/mobile Work needs Plus or above and is rolling out gradually. Free tiers hit limits quickly on long Work.
How much does ChatGPT Work cost?
Plus ~$20/mo; Business ~$20/user/mo annual; Enterprise is quoted (often ~$60/user/mo from 150 seats). Work is not priced separately—it consumes plan Codex/Agent credits with optional add-ons.
How do I install Codex CLI?
Run npm install -g @openai/codex (Node 22+), or brew install --cask codex, or curl -fsSL https://chatgpt.com/codex/install.sh | sh. Launch codex and choose Sign in with ChatGPT.
Do Work and Codex share the same quota?
Yes. OpenAI bills Work usage like Codex. Heavy Work research plus Codex coding on Plus will hit limits sooner; Enterprise spend controls can split usage.
Summary
How to use ChatGPT Work? Install the unified desktop app, switch Work for knowledge tasks and Codex for code, pick GPT-5.6 Sol/Terra/Luna by job type. ChatGPT Work pricing is about credits, not list price—long agents cost more than the monthly subscription. Codex tutorial in one line: install @openai/codex, sign into ChatGPT, run a smoke task in your repo.
Claude Code and Cursor are not either/or—they are entry-point division. Model launches keep accelerating, but engineering teams should fix the execution node (Cloud Mac), connector policy, and credit budget first; let Sol handle the hard problems after that.
Run Work research + Codex coding on Cloud Mac so credits buy delivery
Let ChatGPT Work handle competitor and document work while Codex CLI edits the repo on the same Cloud Mac mini M4—worktree isolation, 24/7 uptime, real xcodebuild validation. Far more stable than a sleeping laptop agent; Apple Silicon unified memory suits Terra/Sol long context, and macOS Gatekeeper + SIP reduce malicious-script risk.
Start with a daily rental to validate Codex smoke tasks and Work connectors, then move to monthly. kvmboot Cloud Mac mini M4 is the cost-effective starting point for ChatGPT Work + Codex—explore plans and spend agent credits on deliverables, not on sleep interruptions.