OpenAI launches ChatGPT Work with Codex and GPT-5.6 desktop agent

• OpenAI shipped ChatGPT Work as a ChatGPT agent powered by Codex and GPT-5.6 that can act across apps and files and stay on projects for hours, according to OpenAI's launch thread.
• The Codex app update becomes the new ChatGPT desktop app: Chat, Work, and Codex are available on macOS and Windows for every plan including Free, while OpenAI Developers' availability post says Sites is in beta for Pro and Plus.
• The Work/Codex split is mostly interface and visibility: dkundel's reply says Work and Codex use the same Codex agent with sandboxing and auto review, while dkundel's follow-up says Work removes developer surfaces such as diffs and PR reviews.
• GPT-5.6 adds Sol, Terra, Luna, Ultra mode, faster Computer Use, a Chrome extension, PR review, Sites, and unified plugins, per reach_vb's launch inventory and OpenAI's GPT-5.6 thread.

The buried gotchas are more useful than the launch name: Simon Willison's screenshot says cloud Work conversations do not appear in desktop Work; the Max-setting PSA says Max reasoning is disabled by default in the new app; jxnlco spotted a unified Plugin Directory; and The Turing Post found a Choose your pet setting with enough companions and effort levels to turn personalization into routing.

OpenAI wrapped the launch around three product surfaces. sama's livestream list named ChatGPT Work, a new ChatGPT desktop app, and hosted Sites.

• ChatGPT Work: OpenAI described Work as an agent powered by Codex and GPT-5.6 that can take action across apps and files, continue for hours, and turn a goal into documents, decks, analyses, sites, and reports in its launch thread.
• Desktop app: Codex app users update into the new ChatGPT desktop app, where Chat, Work, and Codex live together, according to OpenAI's rollout note.
• Developer loop: OpenAI Developers' thread says Codex now supports inline editing in diffs and pull-request review in the side panel.
• Sites: OpenAI Developers' announcement says Sites can build interactive apps with GPT-5.6 and deploy them with hosting, storage, and optional auth built in.
• Rollout: Steven Heidel's rollout post says GPT-5.6 is rolling out across the API, ChatGPT, and Codex, with access expanding over 24 hours.

The cleanest explanation came in replies, not the launch copy. dkundel's reply says ChatGPT Work and Codex are the same Codex agent, with sandboxing and auto review used to reduce risk while preserving the agent's power.

The mode split is presentation:

• dkundel's screenshot shows Work mode as “same power, less technical detail.”
• dkundel's follow-up says Work removes surfaces such as diffs and PR reviews.
• dkundel's UI detail adds that Work hides some shell-command names and does not render diffs.
• Simon Willison's screenshot says Work on web and mobile runs in the cloud, while desktop Work can use local files and apps with permission.
• That same support-page screenshot says cloud Work conversations do not appear in desktop Work at launch, and desktop threads and local files remain on that computer.
• Thibault Sottiaux's memory reply says ChatGPT memory is shared on web and mobile, but desktop memory is still local to the machine.

The new app is also a browser-control launch. jxnlco's browser note says the in-app browser now supports authenticated sites, multiple tabs, file downloads, and persistent tabs.

OpenAI's developer thread lists four pieces of agent infrastructure:

• Computer Use: OpenAI Developers says GPT-5.6 makes Computer Use faster and more token-efficient, with batching and parallel operations across multi-step tasks.
• Supervision: the same thread adds a picture-in-picture view for watching the agent work in the background.
• Chrome extension: OpenAI Developers' Chrome-extension post says the extension can start or resume threads with access to desktop context including local files, projects, plugins, and skills.
• PR workflow: OpenAI Developers' Codex note says more of the coding loop now happens inside Codex through inline diff editing and side-panel PR review.

The PiP detail landed with practitioners because it makes desktop agents less invisible. altryne's PiP post called the computer-use panel a “godsend” because the agent was using different apps while remaining visible.

GPT-5.6 arrives as a three-model family rather than a single model slug. OpenAI's rollout thread says Plus, Pro, Business, and Enterprise users get GPT-5.6 Sol through medium and higher effort settings, with Pro and Enterprise also able to select GPT-5.6 Pro for complex tasks.

The API price ladder is explicit in Adam Goldberg's pricing excerpt:

• Sol: $5 input, $30 output per 1M tokens.
• Terra: $2.50 input, $15 output per 1M tokens.
• Luna: $1 input, $6 output per 1M tokens.
• Cache writes: 1.25x the uncached input rate.
• Cache reads: 90% cached-input discount.

Ultra is the new multi-agent setting. OpenAI's Ultra-mode post says it coordinates multiple agents in parallel, trading higher token use for stronger and faster results on demanding tasks.

The control surface got crowded fast. Sebastian Raschka counted 2 modes, 3 GPT-5.6 models, and 5 effort levels, or 30 possible configurations, and yanndubs replied that there is also fast mode and number of subagents, with consolidation work underway.

Sites moved from demo surface to product surface. OpenAI Developers' availability post says Sites is in beta for Pro and Plus users, with EU and UK availability coming soon and no additional charge during beta beyond plan-specific limits.

The admin details are heavier than the launch demos:

• the release-note screenshot says Business and Enterprise customers can publish Sites publicly through a URL.
• the same release notes say Enterprise public publishing is off by default and must be enabled by an admin.
• btibor91's screenshot says Sites public publishing and expanded beta rollout are not available in the EEA, Switzerland, or the UK at launch.
• the release notes say the Plugin Directory replaces the App Directory, while existing app connections are unaffected.
• jxnlco's plugin post says the directory brings shared discovery, installation, authentication, contextual suggestions, and unified distribution across ChatGPT and Codex.
• Developers can self-serve submissions through the plugin submission page surfaced by jxnlco.

Hands-on reports converged on a specific profile: Sol is fast, persistent, and unusually willing to keep working, while Fable still gets credit for taste and top-end judgment.

• Dan Shipper's day-zero review called GPT-5.6 Sol an A-tier coder, the strongest frontier writer in his testing, and the first model he trusted to run whole loops of knowledge work.
• Shipper listed email processing, meeting and Slack decision tracking, candidate search, Facebook Marketplace scanning, and meal logging in his knowledge-work loop post.
• Matthew Berman's review says Sol ran complex projects for days with /goal, inspected a real Excel app with Computer Use while recreating it, and felt 2 to 3 times faster than Fable in daily use.
• Berman also reported confident mistakes and occasional process leakage in the same review.
• Peter Gostev's comparison called Fable a “wise owl” and Sol a “rottweiler,” with Sol better at diligence, reliability, code-pattern adherence, subagents, computer use, and multi-day runs.
• Ethan Mollick's PDF test says Codex and 5.6 Sol took 30 minutes to produce dozens of notes on a full book PDF, none wrong but many pedantic.
• Mollick's early-access note says Sol works faster in steps, while Fable is better for very long tasks where the user can define the target in advance.

The product merge also upset Codex-heavy users. theo's follow-up framed the launch as a shift away from AI labs focusing almost entirely on software developers and toward office work, spreadsheets, and presentations.

The sharpest critique was about control. Mollick argued that non-coders need more types of control and visibility, not a coding app with options removed.

Other confusion was basic naming and navigation:

• Simon Willison said he was confused by ChatGPT, ChatGPT Codex, ChatGPT Work, Claude, Claude Code, and Claude Cowork.
• Mollick's naming complaint said OpenAI had improved naming with Luna, Terra, Sol, and Codex, then introduced new ChatGPT-based names.
• Ben Hylak's reply asked for Codex back after trying the new ChatGPT Codex app.
• the Reddit AMA screenshot says more than 5 million people use Codex every week, twice as many as three months ago, which explains why a rename could feel larger than a normal app update.

The same release-note bundle included three removals. btibor91's inventory says group chats can no longer be created, converted, or joined by invite link starting July 9; Atlas is scheduled to stop working on August 9; and GPT-5.4 leaves ChatGPT on July 23.

The desktop transition leaves a fallback path. the release-note screenshot says the previous ChatGPT desktop app may remain installed as ChatGPT Classic, with model updates, bug fixes, security patches, and existing Enterprise capabilities, but without the new Work and Codex agent features.