OpenAI announced ChatGPT Work on July 9, 2026, and merged the standalone Codex app into the new ChatGPT desktop app with three modes: Chat, Work, and Codex.
📑Table of Contents
The practical takeaway is simple. Work is not “better chat.” It is a long-running agent that aims for finished deliverables across connected apps and files. Codex remains a coding agent for implementation, tests, and repository work. Treat the launch as a workflow and desktop UX re-layout, not only as another model release.
This guide is for developers and operators already using Codex, Claude Code, or Cursor who need a clear mode split, plan availability map, first-pilot checklist, and permission boundary.
How ChatGPT Work changes the unit of work
Work changes the unit of value from a reply to a deliverable. OpenAI positions it as an agent that gathers context across apps and workflows, breaks complex goals into steps, and produces finished materials such as sheets, slides, docs, and Sites, while the user monitors progress, answers questions, redirects, and approves important actions.
GPT-5.6 shipped the same day and powers Work, but this article keeps model details secondary. The product question is whether you already have multi-hour work units that end in artifacts.
Work-shaped examples include:
- Weekly executive dashboards that pull CRM, email, and docs
- Launch readiness packs that join release plans, trackers, and GTM notes
- Living internal portals or interactive reports that refresh as source data changes
Quick clarification still belongs in Chat. Implementation, tests, and PR review still belong in Codex. Scheduled Tasks can run once, on a schedule, on an event, or while monitoring changes, but they still assume human steering and approval of important actions.
Bottom line: start with the work unit, not the model name.
Chat, Work, and Codex after the desktop merge
Codex is not removed. It merges into the new ChatGPT desktop app on macOS and Windows while keeping a dedicated coding experience. Existing Codex users update as usual, can keep Codex as the default view, and can use the Codex logo as the app icon. The previous desktop client is described as ChatGPT Classic.
A practical mode map:
| Surface | Chat | Work | Codex |
|---|---|---|---|
| Primary job | Q&A, search, brainstorming | Research through finished materials | Implementation, tests, review, repo work |
| Web / mobile | Yes | Paid-plan staged rollout | Not a selectable mode (Remote for some desktop tasks) |
| Desktop | Quick chat and related flows | Local files/apps with permission | Desktop coding mode (all plans, including Free) |
| Usage framing | Normal chat | Same family as Codex; complexity raises consumption | Coding usage pool / extra credits as baseline |
Sources: OpenAI Help: ChatGPT Work and Codex, OpenAI launch post, ChatGPT Learn What’s new (as of July 2026).
Desktop Codex highlights include inline editing inside diffs, PR review in a side panel, faster Computer Use with GPT-5.6, and multi-repository projects. The merge is an entry-point consolidation more than a feature deletion.
Critical launch boundaries:
- Web/mobile Work runs in the cloud
- Desktop Work can use local files and desktop apps when permitted
- Cloud Work conversations do not appear in desktop Work at launch
- Desktop Work threads and local files stay on that machine
- Codex is a desktop mode, not a peer selectable mode on web/mobile
- Mobile Remote can reach supported desktop Codex tasks, but those tasks do not become web/mobile chat history
Chat conversations sync across web and desktop. Work and Codex do not fully share that convenience. Fix the entry surface per project before people fork the same initiative across cloud and local threads.
Plan availability, usage limits, and model choice
Availability is uneven by surface.
- Web/mobile Work starts with Pro, Enterprise, and Edu, then expands to Plus and Business over the following days.
- Desktop Chat / Work / Codex is available on every plan, including Free, globally on Mac and Windows.
- Usage: Work is longer and heavier than chat; it follows the same usage structure family as Codex and rises with task complexity.
- Enterprise: Admin Console spend controls cover workspace defaults, group limits, individual overrides, and credit requests with rationale.
Model choice is the engine, not the whole product. Learn docs currently frame the GPT-5.6 family roughly as:
| Model | Role | Best first use |
|---|---|---|
| Sol | Flagship | Hard coding, computer use, research, security work |
| Terra | Balanced | Everyday work and implementation |
| Luna | Faster / lower cost | Lightweight loops and cost-sensitive tasks |
| Power default | Sol + medium | Safe initial setting when unsure |
Sources: GPT-5.6 release post, ChatGPT Learn What’s new (as of July 2026).
For pilots, measure quality and consumption on Terra or Power first, then escalate only the stuck steps to Sol. Track per-task burn, not only remaining monthly quota.
Pilot checklist and first-task design
Your first win should be one known recurring task, not a platform rewrite.
Routing table
| Question | If yes | If no / unclear |
|---|---|---|
| Do you need a finished doc, sheet, deck, or site? | Work | Chat or human work |
| Do you need code changes, tests, or a PR? | Codex | Do not force Work |
| Is multi-app research required? | Work | Single-turn Chat may be enough |
| Is local repository work central? | Desktop Codex | Cloud Work is a poor fit |
30–90 minute pilot
- Pick one already-understood recurring task.
- Route it to Work or Codex with the table above.
- Connect the minimum plugins and local folders only.
- Write success criteria, forbidden actions, and approval gates first.
- Enable Scheduled Tasks only for supervised routine jobs.
- Run one 30–90 minute pilot and record deliverable quality plus usage burn.
- Save a team template: mode choice, permissions, consumption, failure notes.
Example success criteria
- Work: “One sales weekly dashboard. Leave missing data blank; never invent numbers.”
- Codex: “One PR that turns a failing test green. No unrelated refactors.”
- Approval gate: “Stop before external share links, production config changes, or customer data export.”
Common failure modes
- Vague goals (“make it better”)
- Over-connected apps that expand blast radius
- Ignoring the cloud vs desktop conversation split
- Unattended Scheduled Tasks with no review path
Reader next action, fixed to three steps:
- Choose one known recurring task.
- Decide whether it is Work-shaped or Codex-shaped.
- Run a 30–90 minute pilot with least privilege and approval gates, then record results and usage.
Permissions, security, and operating caveats
Work and Codex expand the surface area of tools, files, and actions. Feature lists are incomplete without an approval model.
Launch-post governance signals include:
- Enterprise foundations for access, company context, connected tools, and allowed actions
- Compliance API visibility into Work conversations and actions
- Desktop controls that build on Codex enterprise governance for local files, apps, browsers, and agent network policy
- Auto-review of important connected-tool/API actions before execution to reduce unintended sensitive sharing
Independent coverage (Ars Technica, Tech Times) adds operational risk the official feature list underplays:
- Work is longer/heavier than chat and can burn more of the same usage family as Codex (“watch your wallet”)
- On the merged desktop, coding agent, built-in browser, Computer Use, and live plugins can sit in one session / trust perimeter
- Earlier cloud Codex emphasized microVM isolation; host files and connected apps expand reach after the merge
- Prompt injection is unlikely to be fully solved, so malicious page/email instructions can fan out across connected systems
- Therefore day-one policy is least privilege for the pilot, not “connect everything”
Even for personal use, review:
- which folders and mailboxes must never be connected
- external send / public Sites / customer-data handling
- who can install plugins
- pilot usage caps and kill switches
The Atlas browser sunset is a migration footnote. The real design problem is boundary control after agents can act across more surfaces.
Safe default: connect only what the first pilot needs, and write the approval gate before the first run.
FAQ
Related articles:
- Codex app 26.609 Update
- OpenAI Codex Record & Replay Lets AI Reuse Screen Workflows
- Building Tabletop Exercises for Incident Response Using Codex as Game Master
Summary
Work is a deliverable agent. Codex is an implementation agent. The desktop merge re-routes entry points rather than deleting coding workflows. Availability and history boundaries differ by surface, so cloud Work and desktop Work/Codex need an explicit operating model.
Do one thing next: pick a known recurring task, route it to Work or Codex, and run a least-privilege 30–90 minute pilot with approval gates. Convert that result into a shared mode-selection template before broader rollout.
Related new article:
- GPT-5.6 Sol vs Terra vs Luna: Pricing, Strengths, and How to Start – This published update adds current operational context for ChatGPT Work Explained: Long-Running Agents and the Codex Desktop Merge.
- GPT-5.6 Sol Token Usage and API Cost Estimation Guide (2026) – This published update adds current operational context for ChatGPT Work Explained: Long-Running Agents and the Codex Desktop Merge.
Author
krona23
Over 20 years in the IT industry, serving as Division Head and CTO at multiple companies running large-scale web services in Japan. Experienced across Windows, iOS, Android, and web development. Currently focused on AI-native transformation. At DevGENT, sharing practical guides on AI code editors, automation tools, and LLMs in three languages.
🔥 Most Popular
- Claude Pricing: I Tested All 5 Plans — Here's My Verdict (2026)
- Claude Desktop Won't Install? Windows & Mac Fixes That Worked (2026)
- How to Spot and Defend Against Two-Stage Phishing Emails in 2026
- Cursor Pricing 2026: Plans & Real Costs After 3 Years of Pro
- Docker Sandboxes (sbx) Guide: Run Claude Code Safely in a microVM
Leave a Reply