OpenAI’s 16 July desktop update puts Chat, Work and Codex into one clearer macOS and Windows experience. The practical change is not merely navigation. It asks users to choose the operating mode before they begin: conversation for a fast answer, Work for an end-to-end deliverable, or Codex for software tasks grounded in repositories and development tools.
The fastest mode is not always the safest or most efficient. A two-minute question does not need an agentic project. A board report with sources should not be scattered across disposable chats. A code change should not be handled as prose when a repository-aware workflow can inspect and verify it.
The mode-selection matrix
| Need | Chat | Work | Codex |
|---|---|---|---|
| Quick explanation or rewrite | Best fit | Excessive | Wrong tool |
| Multi-source report or finished deck | Limited | Best fit | Only if code is central |
| Recurring monitored task | Limited | Best fit where scheduling is available | Useful for software maintenance |
| Repository change and tests | Poor | Possible context gathering | Best fit |
| Brainstorm before committing scope | Best fit | Useful after scope forms | Useful for technical exploration |
| Local file/repository operation | Limited by permissions | Permission-dependent | Repository and tool focused |
OpenAI’s ChatGPT release notes say the refreshed desktop layout is live for all plans on macOS and Windows, with unified Recents, Projects in the app and cloud Work conversations that continue across devices. The same notes distinguish local conversations, which stay on the computer, from cloud Work conversations that sync. Review OpenAI’s data controls guidance and your workspace policy before treating the interface as a universal storage location.
Use Chat when the output is the conversation
Chat is the correct starting point for a question, a short comparison, a draft paragraph or a decision you want to think through interactively. The conversation itself is the workspace. It is easy to redirect and low-cost to abandon.
Keep the prompt bounded. Provide the facts required for the answer, not an entire confidential archive. If the discussion becomes a multi-step deliverable with files, sources and review criteria, move it into a Project or Work workflow rather than repeatedly pasting context.
Use Work when the output is an artefact
Work is designed for tasks that have a beginning, checkpoints and a finished object: research reports, spreadsheets, presentations, sites or monitored processes. Start with an acceptance brief: audience, required sources, constraints, output format, date boundary and actions that require approval.
The key discipline is permission design. Connected apps and local files can make an agent useful, but every added source expands what it can see and what an adversarial document may try to influence. Grant the narrowest practical access, review proposed external actions and remove connections that are no longer needed.
Use Codex when correctness lives in the repository
Codex is the specialist choice when the answer must become a code diff, test result or repository operation. It can inspect project instructions, make scoped edits and verify behaviour with development tools. A coding request in general Chat can produce plausible snippets; a repository-aware task can discover that the snippet conflicts with the actual architecture.
Define the working directory, branch expectations, files that must not change and the verification command. For production data or deployment actions, require a preview and explicit evidence. A successful-looking diff is not a live verification.
Projects and Recents change retrieval, not governance
Unified Recents and Projects reduce the cost of finding work. They do not remove the need for names, ownership and retention rules. Use a consistent project name, place source files in the correct boundary and archive obsolete experiments. Pin only active work so the sidebar remains a control surface rather than a second inbox.
When continuing across devices, verify whether the conversation is cloud-based or local. Do not assume a locally grounded workflow will appear elsewhere, and do not move sensitive local context into a cloud task merely for convenience.
A 60-second routing decision
Ask four questions before starting:
Is the desired output an answer, an artefact or a code change?
Does the task require files, apps, a repository or the open web?
Which actions could change external state?
Where should the context live after the task ends?
Answer, no tools: Chat. Artefact with multiple sources and review: Work. Repository change with tests: Codex. If the task crosses modes, start in Chat to shape the brief, then hand the stable brief to Work or Codex.
The broader AI agent and human concierge disruption matrix applies the same boundary: automation is strongest when it owns repeatable preparation and a human retains consequential judgement.
A migration check for existing desktop users
Before removing an older app, identify which conversations are local, which are cloud-synchronised and which files exist only on the computer. OpenAI says the previous app may remain as ChatGPT Classic and local conversations stay on the machine. Do not assume installation copied every local artefact.
Inventory pinned projects, connected apps, permission settings and scheduled work. Test one low-risk task in each mode, then verify where its history appears on another device. Enterprise users should follow administrator guidance on retention and connectors.
A worked task from question to verified output
Consider an executive team preparing a market-entry briefing. The first question—“What changed in the target market this quarter?”—belongs in Chat. The user can refine the scope, identify ambiguous terms and decide which claims require current primary sources. At this stage, speed and dialogue are more valuable than a polished deliverable.
Once the brief is stable, move the work into an artefact-oriented workflow. Define the date range, approved sources, required charts, audience, length and review owner. Work can gather the evidence and produce the report or presentation, while the human checks that citations support the displayed claims. The output is not complete because a file exists; it is complete when the acceptance criteria are met.
If the briefing also requires changes to a data pipeline or dashboard, that portion belongs in Codex. Give it the repository, data contract, tests and deployment boundary. Keep the market judgement in the report workflow and the software correctness in the repository workflow. Crossing modes is acceptable when the handoff is explicit and each stage retains the appropriate evidence.
Match permission to the mode
The three modes should not automatically inherit identical access. A quick Chat rarely needs a connected drive. A Work task may need a particular folder but not the ability to send the final document. Codex may need repository write access and tests but no access to customer email.
Use a permission worksheet:
| Question | Chat | Work | Codex |
|---|---|---|---|
| Minimum context | Facts needed for the answer | Brief, sources and output files | Repository and task specification |
| Typical write surface | Conversation only | Named artefacts or approved apps | Scoped files and branch |
| External action boundary | Ask before messaging or publishing | Preview and approve delivery | Separate approval for deploy or production data |
| Verification | Check reasoning and current sources | Review artefact and citations | Run tests and inspect diff/live behaviour |
| Retention review | Delete disposable sensitive context | Archive project under policy | Preserve change history in repository |
Review permissions when the task changes. A conversation that evolves into a deliverable does not gain authority merely because the interface makes the transition convenient.
Know when not to use the larger mode
Agentic work carries coordination cost. A short rewrite can take longer if the system creates a project, searches several sources and produces an artefact that must be reviewed. Use the smallest mode that can satisfy the acceptance test.
Avoid Work when the answer is subjective exploration with no stable output yet. Avoid Codex when there is no repository or executable verification. Avoid Chat for a multi-file deliverable whose evidence and revisions need to remain organised. The boundary is the source of correctness: dialogue, artefact or codebase.
Teams should measure correction time, not only generation time. If a mode repeatedly produces impressive but mis-scoped output, improve the routing brief before adding more tools. The most capable interface cannot compensate for an undefined owner, missing source boundary or absent acceptance test.
Add one final routing check for collaboration. If several people must review the same evolving evidence, keep the work in a shared, governed project rather than forwarding screenshots from private chats. If only a developer can verify the result, move the technical portion into the repository and return the tested outcome to the shared artefact. The place where colleagues can inspect and approve the work is part of choosing the mode, not an administrative detail added at the end.
The important update
Teams should publish a one-page routing rule rather than leaving every user to infer the difference. Name the approved mode for quick questions, finished internal artefacts and repository work; identify which connectors are allowed; and state who reviews external publication. Revisit the rule when plan availability or workspace controls change. A shared operating definition prevents the same sensitive task from being handled three different ways simply because each person opened a different tab.
The July desktop release makes three AI working styles easier to reach. Its value comes from choosing deliberately. Chat is not an inferior Work; Work is not a bigger Chat; Codex is not a general-purpose label for any difficult task. Each mode reduces a different kind of friction—and introduces a different permission and verification burden.
Route the task before supplying the context. That one habit will matter longer than the current layout.
