OpenAI renames Codex Superapp to ChatGPT Work in desktop app

• OpenAI rolled Codex into the ChatGPT desktop flow: the install dialog in thekitze's screenshot says “Codex is now the ChatGPT app,” while LLMJunky's screenshot shows the new ChatGPT Work surface.
• ChatGPT Work is built around tasks, plugins, scheduled runs, sites, and chat, with file-system and computer-use behavior described in LLMJunky's explanation as closer to Codex than the old ChatGPT desktop wrapper.
• The first rollout bugs were visible fast: the Work/Codex sidebar stayed the same in thekitze's toggle test, Sites returned a 404 in petergyang's publish test, and credits failed for bentossell's token complaint.
• GPT-5.6 Sol is part of the day-zero story: danshipper's month-long test called it the best writer among frontier models and a knowledge-work leap, while still putting Fable ahead on high-end coding.
• Creators immediately turned the app into a workflow machine: gregisenberg's masterclass thread listed inbox cards, Slack and meeting feeds, agent email, watch-once skills, and 20-hour runs.

The rebrand dialog keeps one foot in the old world: “Keep coding with Codex,” “Work beyond code,” and a checkbox to keep the Codex icon. The first ChatGPT Work screenshot already has Sites, Plugins, Scheduled, Chat, Projects, and a prompt asking the agent to organize a messy Notes app. The full Every Vibe Check focuses on writing and knowledge loops, while petergyang's live design walkthrough shows GPT-5.6 building a travel site with a 3D Torii gate.

The rollout text in the installer said “Codex is now the ChatGPT app,” then split the pitch into two promises: “Keep coding with Codex” and “Work beyond code.” The checkbox at the bottom let users keep the Codex app icon.

LLMJunky called the product the former “Codex Superapp” now named ChatGPT Work in his first screenshot, then posted that users could preserve the Codex branding through the icon and alias settings in his follow-up. When asked how to get it, LLMJunky's update reply said it arrived as an update.

The product move is obvious, the naming hit like a wet conference badge. LLMJunky still argued in one reply that the strategy made sense because most people know OpenAI as ChatGPT, even though Codex had become one of the company's better names.

The most repeated complaint was mode confusion. In thekitze's test, switching between ChatGPT Codex and ChatGPT Work changed nothing visible in the sidebar.

The dropdown in thekitze's screenshot labels Work as “For getting work done” and Codex as “For developers.” That wording prompted danshipper's joke that it implies developers do not do work.

petergyang's longer feedback thread said the split raised basic product questions: vacation planning could be “work,” model choices like Sol, Terra, and Luna had no in-UI explanation, and old chat history felt harder to find in the new task-first structure. petergyang's feedback thread also said GPT Live may matter more for mainstream users than the desktop agent shell.

The harshest proposed fix came from thekitze's two-app argument: keep Codex as the developer studio, keep ChatGPT desktop for chat and general work, and stop mixing the two jobs into one ambiguous switcher.

The first Work surface opens with “What should we get done?” instead of a blank chat prompt. The visible sidebar in LLMJunky's screenshot lists:

• New task
• Scheduled
• Plugins
• Sites
• Chat
• Pinned
• Projects
• Tasks

The sample task list is ordinary office work: visualize feedback themes, compare travel headphones, prepare for a manager 1:1, review a resume, plan a Seattle weekend, draft a weekly team update, create a spreadsheet visualization, and prepare launch readiness.

LLMJunky described the old ChatGPT app as closer to the web interface in a desktop wrapper. ChatGPT Work, according to his explanation, is closer to Codex: a coding agent and general work agent that can act on the computer's file system.

A Turkish walkthrough from ozansihay framed the same surface as a finished-work layer: documents, tables, presentations, reports, and shareable sites, with Chat, Work, and Codex living together on macOS and Windows.

Sites was present in the sidebar before it was smooth in practice. petergyang's test showed a 404 page, then a task response saying the site was validated and ready but Sites was not connected to the account.

He later told another user he would deploy the project “when OpenAI fixes the sites feature” in his follow-up, which also linked to a GPT-5.6 versus Fable travel-site comparison.

bentossell hit a separate account-level snag: his token post said he was out of tokens and could not add more credits. A later mobile screenshot from bentossell showed Codex surfacing as “Remote” inside the ChatGPT app menu.

GPT-5.6 Sol arrived in the same social burst as the desktop app merge. The most useful hands-on split came from danshipper's month-long test, which compared it against Fable and Opus 4.8 across coding, writing, design, and knowledge work.

His findings were compact:

• Coding: Sol scored 56/100 on an internal Senior Engineer benchmark, versus 91 for Fable.
• Writing: he called it clearer and more concise than Fable or Opus 4.8.
• Design: he saw an improvement over GPT-5.5, while still putting Fable and Opus 4.8 ahead.
• Knowledge work: he trusted it for whole loops across email, Slack, meetings, candidates, Facebook Marketplace, and meal logging.
• App shell: he said Work and Codex are “basically the same” except Work hides code.

petergyang's first-day read was warmer on persistence and colder on product shape. In his feedback thread, he said GPT-5.6 “basically never gives up,” seemed reliable with browser use, burned more tokens than GPT-5.5 on Sol High, and left him confused about when to use Sol, Terra, Luna, or different effort levels.

Greg Isenberg's launch thread cited OpenAI benchmark claims for Sol at 88.8% on Terminal-Bench 2.1 and 91.9% in ultra mode. Those numbers appear in gregisenberg's thread as benchmark claims, not an independent test.

The creative demos were less about code completion and more about agent routines. Greg Isenberg's thread, built around a 49-minute masterclass, listed five workflows:

1. Turn the inbox into daily cards with summaries and replies drafted in the user's voice.
2. Convert Slack, meeting notes, and company updates into a daily feed with next actions.
3. Give the agent its own email address so tools and team bots can route work to it.
4. Watch a human do a task once, then turn the observed process into a repeatable skill.
5. Run a long goal for 20 hours, including fine-tuning a model.

petergyang's design test used GPT-5.6 and Fable 5 to build travel sites for a Japan trip. In his side-by-side demo, GPT-5.6 made a site with a 3D Torii gate and he gave the prompt trick: ask for “a 3D WebGL element to the hero.”

MengTo said he was “pretty much all in on Codex” after GPT-5.6 Sol. his breakdown post named agent skills, goals, spawned threads, browser use, computer use, and a mobile connector, then said he used Codex to ship four major apps and organize video and written content.

Aakash Gupta's Meng To breakdown reads like a creative-agent syllabus: Codex setup, AI browser, real project folders, plugins versus skills versus computer use, when to grant full access, screenshot shortcuts, a senior designer taste skill, and a digital twin. aakashgupta's chapter list puts the workflow details in timestamps instead of vague hype.

OpenAI's own icon also got a maker note: Gavmn's Blossom post said the ChatGPT Mac app icon got a small touch-up, while his Blender follow-up showed the logo being lit and rendered before final compositing.

The app now has enough parallel agent work that thread management becomes product design. LLMJunky complained in his Codex UI feedback that active tasks were identified by tiny spinners that were easy to miss when scanning dozens of threads.

His proposed fix was visual state: highlight active and completed threads more clearly, with customizable colors or conditions. A reply in LLMJunky's follow-up suggested using those highlights for active thread automations too.

The same user found a housekeeping use case for the agent itself. In LLMJunky's stale-thread cleanup, Codex archived repeated automation tasks and verified that 30 matching tasks no longer remained active.

A lower-level computer-use detail came from steipete, who posted the macOS defaults command ComputerUseAllowForbiddenTargets in his command post. In his security reply, he said it fixed codex-on-codex behavior and was “obviously also a potential security issue.”

The desktop merge also showed up on mobile. bentossell's screenshot displayed a ChatGPT app menu with Library, Projects, Scheduled, Plugins, Remote, and Images, with Remote marked by a computer icon and green status dot.

Rohan, identified as a Codex PM in petergyang's calendar clip, described using Codex as a personal chief of staff: review the week's meetings, consolidate focus blocks, and identify meetings that could become DMs. The full episode link sits behind the clip for the longer version.