ChatGPT Work and a new desktop app: OpenAI merges ChatGPT, Codex and the browser

What's going on

OpenAI is overhauling its ChatGPT products. At the center is ChatGPT Work, an AI agent: a program that works through multi-step tasks on its own instead of just answering questions. According to Vietnamese newspaper Báo Cần Thơ, citing Neowin, it runs on the GPT-5.6 model family. OpenAI has also released a completely new ChatGPT desktop app for Windows and macOS that includes the Codex coding agent and a web browser. On top of that, there's a beta of the Sites website generator, a new Chrome extension, and the end of Atlas, OpenAI's own AI browser.

According to ZDNet, Betanews and Cryptobriefing, the separate "Chat" and "Work" tabs will be merged next. Starting in December, ChatGPT and Codex will also share a single usage allowance.

Work is available first to ChatGPT Pro, Enterprise and Edu users on the web and in the mobile apps. Plus and Business customers should get it over the next few days. The new desktop app is free worldwide.

What ChatGPT Work can do

Work is designed to operate across apps, files and workflows. The agent pulls data from different sources and turns it into finished output: spreadsheets, presentations, documents, reports and even web apps. It follows any templates and reference documents you give it.

Data sources connect through plugins, extensions that give ChatGPT access to other services. OpenAI lists Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRM systems (customer management software) and project management tools. Type "@" in your prompt to have ChatGPT pull data or context from a specific app.

Also new is Sites, launching as a public beta. It turns almost any data source into a website or interactive web app without any coding. Think dashboards, project trackers, product roadmaps, internal portals or visual reports. When the source data changes, these pages are supposed to update automatically.

OpenAI isn't alone here. Anthropic, Microsoft and Google already offer similar agents to business customers. According to ZDNet, Anthropic merged Claude's chat and Cowork features into a single interface in September.

The new desktop app: Codex, browser and Computer Use

The new desktop app now includes Codex, which gives ChatGPT access to local files and programs on your machine. A built-in browser handles online tasks right inside the app. On Windows and macOS, you also get Computer Use: the AI operates files, programs, tools and websites itself, much like a person with a mouse and keyboard.

Codex is still a separate, specialized coding agent with its own section in the shared interface. If you had the Codex app installed, you'll be moved to the new ChatGPT desktop app automatically. Developers can make Codex their start view and even pick the Codex logo as the app icon. The previous ChatGPT desktop app is now called ChatGPT Classic.

In the browser, a new Chrome extension opens ChatGPT in the sidebar. OpenAI will no longer develop Atlas as a standalone product.

Chat and Work are merging

The desktop app is a bit confusing right now, as ZDNet also points out. There's a ChatGPT section and a Codex section, and the ChatGPT section has its own Chat and Work tabs. That's why OpenAI plans to fold Work's agent features into the regular chat. An email to subscribers, reported by TechCrunch, ZDNet and Betanews, says:

"Go from first idea to ChatGPT getting it done, all in one place. You can bring in your files and plugins and ask ChatGPT to build a website, create a presentation or get things done on the web for you, while you review and steer its work."

A select test group gets the merged interface first and will provide feedback. According to Betanews, the email doesn't give a date for the wider rollout. Codex stays a separate section in the desktop app.

Shared token allowance from December

For paying users, billing is the biggest change. Starting in December, ChatGPT and Codex will share a single allowance on the Plus, Pro and Business plans. It's measured in tokens, the chunks of text a model reads and writes. How much you use depends on the model and the task. Stronger models and heavier tasks burn through more. "Instant chat doesn't count against your allowance," OpenAI writes. You can check your remaining balance and reset times under Settings > Usage. Once you hit the limit, you can wait for the reset, upgrade to a pricier plan or buy credits.

OpenAI expects most users won't have to wait for a reset any more often than they do today. According to Cryptobriefing, the plans still follow different rules: Plus and Business have a 5-hour limit, while Pro doesn't but may have weekly caps. How those caps will work with the shared pool is still unclear.

The criticism has been loud. According to Betanews, the story got more than 220 points and over 160 comments on Hacker News. Many people complain that usage is billed in tokens even though no plan says how many tokens it includes. Sure enough, neither the email nor the help article gives actual numbers. Cryptobriefing also warns that your own usage will now compete with itself: the more you code with Codex, the less you'll have left for regular chat.

The limits have also gone back and forth. On September 29, OpenAI announced it would cut the agent and coding allowance on the Pro 200 plan in half starting October 30, with no price cut. On October 29, executive Thibault Sottiaux ("Tibo") reversed course on X:

"Starting today, we're doubling all usage limits for all Pro 100/200/500 plans and Business Premium. We heard your kind feedback and won't be lowering the limits for Pro 200."

Plus allowances stayed the same. Only Pro 500, at $500 a month, offers "Ultrafast" versions of GPT-6 Astra and GPT-6.1 Sol. OpenAI itself notes that double the allowance doesn't mean double the work gets done.

Our take

OpenAI is cleaning up its messy product lineup and betting on agents that handle office work from start to finish. That brings it level with what Anthropic already offers with Claude, and what Microsoft and Google offer with their business agents. For users, the most concrete improvements are the free desktop app with Computer Use and the deeper hooks into Slack, Teams and Google Drive. Killing Atlas shows that OpenAI no longer sees a standalone browser as worth pursuing.

Billing is the weak spot. A shared token pool with no published amounts makes it hard to plan your usage. And the repeated reversals on limits show the numbers could change again at any time. The model details are murky too: the main report says Work runs on GPT-5.6, while the plan information talks about GPT-6 Astra, GPT-6 Pro and GPT-6.1 Sol. Which models Work actually uses day to day is still an open question.