CHATGPT
WORK_
CODEX_
MERGE_2026.

Тезис: 9 июля 2026 OpenAI выпустил GPT-5.6 (Sol / Terra / Luna) и одновременно провёл архитектурный merge: автономное Codex desktop app EOL, весь stack перенесён в новое ChatGPT desktop application; параллельно — launch ChatGPT Work, work-agent с multi-hour autonomy и deliverable output. Ниже — полный technical breakdown: Codex merge path, tri-mode architecture (Chat/Work/Codex), 1400+ plugin catalog, Plan Mode gate, Computer Use surface, Scheduled Tasks, Codex post-merge upgrades, head-to-head с Claude Cowork, pricing model, 5-step setup, FAQ и Mac/Metal angle.

1. Три engineering-вопроса после релиза

1. «Codex убили?» — Standalone app deprecated, но Codex mode в desktop app сохранён и получил Diff inline edit, PR sidebar, faster Computer Use (GPT-5.6 backend). Existing users: in-place update, zero reinstall.
2. «Work vs Cowork — где execution boundary?» — Work: cloud plugin orchestration (Slack/Gmail/Salesforce API calls через unified catalog). Cowork: local folder sandbox + native M365 add-ins. Не XOR — большинство prod teams будут run оба pipeline параллельно.
3. «Usage billing spike?» — Work и Codex share единый usage pool. Task complexity × runtime = consumption. OpenAI не публикует unit economics — baseline через Plan Mode pilot обязателен.

2. Что произошло 9 июля 2026

2.1 Codex standalone → EOL

С 9 июля 2026 автономное Codex desktop app снято с дистрибуции. Full capability set migrated в новое ChatGPT desktop app. Installed Codex app → auto-upgrade через update channel. Projects, settings, workflow state preserved. Legacy ChatGPT desktop renamed ChatGPT Classic — still available.

2.2 Tri-mode architecture

Mode	Execution profile	Target user
Chat	Standard dialog/Q&A	All tiers
Work	Multi-hour cross-app agent, deliverable output	Knowledge workers
Codex	Code review, PR management, multi-repo ops	Dev teams

Все тарифы включая Free — tri-mode доступен через desktop (Free rate-limited).

2.3 ChatGPT Work: execution model

Work agent pipeline:

1. Plan Mode — step enumeration, user approval gate
2. Plugin auth + context fetch (Slack, Gmail, Google Drive, …)
3. Multi-step autonomous execution
4. Deliverable: doc/spreadsheet/deck/web app

Backend inference: GPT-5.6 family powers Work + Codex Computer Use latency improvements.

3. ChatGPT Work: feature stack

3.1 Plugin catalog: 1400+ connectors

• Collab: Slack, Microsoft Teams, Zoom
• Storage: Google Drive, SharePoint, Dropbox
• Mail/Calendar: Gmail, Outlook, CRM calendars
• Sales/Marketing: Salesforce, HubSpot, LinkedIn, Adobe
• Dev: GitHub, Canva, Zapier

Invocation: @app_name в prompt или natural language task description — Work resolves data source automatically.

3.2 Plan Mode — approval gate перед execution

Complex tasks → step plan → explicit user confirm → run. Prevents agent drift, создаёт audit checkpoints для compliance-sensitive workflows.

3.3 Computer Use — desktop automation surface

• Local file read/write
• Built-in multi-tab browser (web research/scraping)
• UI automation: click, type, file ops
• One-shot или Scheduled Tasks (cron-like triggers)

Dedicated browser Atlas → sunset phase; Work built-in browser + Computer Use absorb capability.

3.4 Deliverables, не drafts

• Docs (Word/PDF reports, email drafts)
• Spreadsheets (Excel/Sheets, financial models)
• Decks (Slides/PPT with template injection)
• Web Apps (Codex Sites interactive output)

3.5 Scheduled Tasks — background agent loop

Time-triggered, event-triggered или periodic execution — agent runs без active user session at keyboard.

4. Codex post-merge upgrades

• ✅ Diff inline editing — edit directly in diff view
• ✅ PR sidebar review — zero context switch
• ✅ Faster Computer Use — GPT-5.6 inference path
• ✅ Multi-repo project scope — single project, multiple repos

Dev config: Codex as default launch mode; macOS — retain Codex dock icon. Desktop Codex projects accessible from mobile ChatGPT app (remote monitor/control).

5. ChatGPT Work vs Claude Cowork — comparison matrix

Dimension	ChatGPT Work	Claude Cowork
Runtime	Cloud + desktop hybrid	Local desktop-first
File I/O	Desktop: local FS; web: upload	Sandboxed local folder
Integration	1400+ plugins	20+ MCP connectors, native M365
Best fit	Cross-web/SaaS orchestration	File-heavy doc pipelines
Non-tech UX	High	Very high
Pricing	Usage-based	Seat-based (Pro $20/mo)
Free tier	✅ Desktop (limited)	❌
M365 native add-ins	Web-only	✅ Word/Excel/PPT

Routing: SaaS/browser workflows → Work; local file batch ops → Cowork. Prod teams: dual-stack.

6. Setup — 5 steps

6.1 Desktop (recommended, all tiers)

1. chatgpt.com/download — new ChatGPT desktop (Mac/Windows)
2. Existing Codex install → update in place
3. Switch to Work mode (top bar)
4. Connect plugins (Gmail, Slack, Drive…)
5. Describe task → review Plan Mode → confirm execute

6.2 Web rollout

• Pro/Enterprise/Edu: live 9 July 2026
• Plus/Business: rolling days after
• Free: web restricted — use desktop

6.3 Mobile

Pro/Enterprise/Edu first; Plus/Business follow. Mobile = monitor/manage; heavy compute stays on desktop host.

7. Pricing & usage pool

Work не отдельный SKU — included in existing plans. Usage-based billing shared with Codex. Complexity × duration = consumption. Enterprise Admin Console: workspace/group/user spend caps.

Plan	US$/mo	Work access
Free	$0	Desktop limited
Go	$8	Desktop extended
Plus	$20	Desktop + web/mobile
Pro	$100–$200	Full, max quota
Business/Enterprise	Custom	Full + Admin Console

8. Strategic read: super-app consolidation

1. Model race → workflow depth race — GPT-5.6 matters; strategic bet = embed in Slack/Drive/desktop.
2. Codex audience expansion — 5M weekly users, 1M+ non-dev workloads.
3. Super-app shape — Chat + Agent + Coding + Computer Use + Scheduled Tasks + 1400 plugins. Andrew Ambrosino (Codex lead): merge = «first step only».

Product thesis: deliverable-output AI beats advisory-only AI.

9. Mac / Apple Silicon: operational signals

• Desktop = agent primary runtime — Computer Use, local FS I/O, multi-tab browser = desktop-only surface. Web-only users miss core Work capabilities.
• Shared usage pool Codex + Work — parallel long-running agents on MacBook saturate Apple Silicon unified memory; battery drain + foreground contention under multi-hour Scheduled Tasks.
• Metal / Apple toolchain — Work Computer Use handles generic desktop automation, но Metal-accelerated pipelines (FCP, Resolve, Blender Cycles Metal) и native Apple dev toolchain (Xcode, Instruments) требуют bare-metal macOS — Windows cloud VM не substitute.
• Cowork complement — local folder sandbox + M365 native: Cowork wins; Slack/Gmail/Salesforce cloud orchestration: Work wins. Typical dual-stack: Cowork batch-summarize local contract repo; Work pull Salesforce → generate weekly report → Slack notify.

Confirmed by OpenAI Blog, The Verge, MacRumors, Ars Technica — global Mac/Windows rollout 9 July 2026 alongside GPT-5.6 GA.

10. FAQ

Q1: Codex standalone app still works?

No. Merged 9 July 2026. Update in place; projects intact.

Q2: Free users get Work?

Desktop yes (limited). Web/mobile restricted.

Q3: Work vs Agent mode?

Work = long cross-app deliverables + Plan Mode. Agent = short single-step.

Q4: Cost unpredictable?

Usage-based; no published unit price. Plan Mode pilot first.

Q5: Codex projects lost?

No. Preserved; mobile accessible.

Q6: Work vs Operator?

Operator = browser automation only. Work = plugins + local FS + multi-hour projects.

Q7: Switch from Cowork to Work?

Workflow-dependent. Most teams run both.

11. Sources

• OpenAI Blog: ChatGPT for your most ambitious work
• The Verge: GPT-5.6 & ChatGPT Work
• ChatGPT Learn Changelog
• MacRumors: ChatGPT Work launch
• Digital Applied: ChatGPT Work agent launch

12. Conclusion: local Mac for control, remote node for long agent runs

Codex merge positions Mac as OpenAI super-app entry point — Computer Use + local FS + 1400 cloud plugins handle short interactive tasks on MacBook fine. But multi-hour autonomous agents, multi-repo Codex projects + Work Scheduled Tasks in parallel → unified memory pressure, thermal throttling, battery collapse; pure Windows cloud host lacks Apple toolchain + Metal GPU path for creative/dev side workloads.

Pragmatic architecture: local Mac = Work/Codex UI + light tasks; long-running agents + 7×24 Scheduled Tasks → MACGPU remote Mac mini M4 node — Apple Silicon unified memory for parallel agent processes, SSH on-demand, «foreground approve + background execute» split. Keeps Metal/FCP/Resolve/Xcode stack intact while offloading agent memory hog from laptop.