ChatGPT Work, OpenAI's agentic environment at chatgpt.com/work, changes what a ChatGPT session produces. Instead of returning text in reply to a question, it accepts a goal and a set of connected tools, then plans the job, browses the web, operates software, hands pieces off to background helpers and returns a finished deliverable such as a working website, a slide deck or a set of executed trades. Three hands-on cases, a corporate launch party, a family theme park trip and a self-refreshing news dashboard, show how far that shift has gone and what it may mean for people who use software and people who build it.
Chat, Work and Codex: three modes, three kinds of output
OpenAI merged ChatGPT, Codex and the Work capabilities into one interface roughly six to eight weeks ago. ChatGPT Work is powered by frontier models such as GPT-5 Astra and is positioned as the layer that turns scattered ideas into completed work. An official comparison matrix separates the three modes by purpose, input and result.
| Mode | Purpose | Typical input | Result |
|---|---|---|---|
| ChatGPT Chat | Answer questions, explain concepts, generate ideas | Conversational prompts | Text |
| ChatGPT Work | Complete whole tasks autonomously | High-level goals and project files | Finished artifacts such as sites, decks and files |
| ChatGPT Codex | Development work inside a codebase | Repository and feature or bug descriptions | Working code and commits |
The practical difference is that Work interacts with the computer directly. It opens a browser, handles authentication prompts, manipulates local applications and runs multi-step tasks in the background. A simple example is asking it to clear an entire afternoon of calendar commitments, something that previously meant manual configuration. Just as important, the environment hides the code entirely. Earlier tools such as Codex and Claude Code showed raw code, which intimidated non-technical users. Here, a business user can describe a tool and get a working one without ever reading a line of it.
Case 1: Planning a product launch party with a live project site
The first case comes from a software company preparing a launch party in San Francisco for a new AI workflow product. Inside ChatGPT Work, the user created a project named for the event and entered a prompt asking the agent to plan and execute the project through an internal website: an event on October 20, 2026, for roughly 100 guests.
ChatGPT Work spun up subagents, which are smaller background agents assigned to specific subtasks, and used its Sites feature to generate an internal web application with tabs for Overview, Venues, Tasks, Budget, Guests and Run of show. The agent browsed the web for venues near the Embarcadero Center and pulled capacities, standing-room figures and photos into side-by-side cards with quote requests and walking routes. The whole research-and-build cycle took about 5 to 10 minutes.
| Venue | Listed capacity |
|---|---|
| Sens - Biada Lounge, Embarcadero Center | 130 standing |
| Bar Sprezzatura | 200 standing indoors |
| Heartwood, Commercial Street | 120 guests |
| The Treasury, Bush and Sansome | Up to 200 guests |
| Harrington's Bar & Grill, Front Street | 200 guests |
| Schroeder's, Front Street | 300 standing |

▲ Agent-built launch party planning site
The finished dashboard showed a 36-day countdown, a task counter at 0 of 16, confirmed guests at 0 of 100 and a planned budget of $0, ready to fill in as work progressed. When the user asked to bring the company's branding into the site, ChatGPT created two dedicated subagents, one for brand assets and one for venue photos. The photo agent searched for each venue, downloaded high-resolution images and embedded them on the page. If the shortlist is wrong, the user can ask for new research and the live page updates without restarting the conversation.
The argument for this approach is organizational: a persistent internal site gives team leads and executives real-time visibility into milestones, which may remove the need for status meetings.
Case 2: A family trip with split routes and a slideshow for a four-year-old
The second project, a four-day summer trip to Disneyland, used family constraints as the prompt: older boys who want roller coasters and a four-year-old daughter who needs gentle attractions. ChatGPT produced an itinerary site with separate routes, split schedules, checklists, daily notes, meal recommendations and Google Maps routes through the park. Each parent can view the plan for whichever group they are escorting.
The same research then became a second artifact. A companion slide presentation, made for the daughter to preview rides and give feedback, featured Peter Pan's Flight, Dumbo the Flying Elephant, Storybook Land Canal Boats and It's a small world with photos and simple captions. The point is that data gathered once in a chat can be reused across websites, decks and spreadsheets.
An earlier trip to Disney World in Orlando tested a ChatGPT-built itinerary in the real world. Without being asked, the agent embedded clickable Google Maps links, so tapping an entry on a phone opened walking directions between attractions, starting with the Hulk coaster. Compared with the single printed sheet of a traditional agenda, a responsive page on a smartphone can be adjusted and shared on the fly.
Case 3: A news dashboard built and run from a phone
The third case shows the agent as an ongoing service rather than a one-off builder. An AI news dashboard, hosted on ChatGPT Sites, solves the problem of tracking fragmented sources across social media, blogs, YouTube and developer releases. A scheduled ChatGPT task runs every six hours, scans leading AI creators and official feeds from OpenAI, Anthropic, Google and GitHub, compiles about 30 top stories, categorizes and summarizes them and assigns each an importance score. In one refresh, a GPT-6 Astra item scored 89, a Claude Cowork item 86, Gemini 3.8 Flash 84 and DeepSeek V4.1 Flash 79. Each card links to the source, official announcement, GitHub repository or an embedded YouTube breakdown, and filter tabs cover companies such as OpenAI, Claude, Gemini, Perplexity, DeepSeek and Hermes.

▲ Always-on home machine powering a mobile news dashboard
The infrastructure behind it is an always-on Mac mini at home, connected to ChatGPT through scheduled tasks and remote access tools. That setup lets the owner conceive an app during a walk, direct ChatGPT from a phone to build and deploy it, and check the mobile-optimized result in Safari on an iPhone. ChatGPT Sites handles responsive layouts, custom domain names and automatic background refreshes, so there is no separate hosting to configure. The same machine runs a multi-channel content engine producing newsletters, carousel PDFs and social copy, plus a creator business tracker. The earlier wave of enthusiasm for home Mac minis in the OpenClaw era appears, in this reading, to have been a sound idea that was pointed in the wrong direction at first but fits persistent remote agents well.
Keeping long runs on target
Several techniques make these long background jobs reliable.
- Start with a goal specification. Typing /goal rather than /plan makes the model write a full spec sheet and a definition of the end deliverable, for example an implementation document. Over runs that last one to two hours, the agent keeps checking its work against that document before declaring the job done. The failure this prevents is what might be called agent "ADHD": wandering into side tasks and forgetting the original request as the context window gets compacted. Newer ChatGPT models appear to understand goal-setting from plain language, so the explicit slash command is often optional.
- Use Schedules for overnight work. A plain-text request to keep iterating on a project until morning ran without intervention, where other tools reportedly needed a /loop command every 15 minutes until 9 AM.
- Bridge to outside APIs with MCP. Model Context Protocol, or MCP, is a standard for connecting an AI model to external tools. A basic MCP server on the Mac mini links ChatGPT to the YouTube API and web search, letting it analyze hundreds of high-performing thumbnails from reference creators without copying them. Custom Skills store design guidelines and brand assets, and the official Google Drive connector saves the output.
- Use native image generation. ChatGPT 2.5 Image Gen produces thumbnails, LinkedIn carousel PDFs and icon suites inside the same workspace, and a Skill instructs the model to fetch exact vector logos and brand guidelines rather than approximate them.
What this means for daily life and the software business
The system diagram for ChatGPT Work lists its modules: Projects for sandboxed research, Plugins for enterprise tool connections, Browser Use and Computer Use for autonomous execution, Schedules for recurring routines, Blocks for reusable text assets and Skills for persistent instructions. Taken together, these overlap with what many standalone AI startups sell, including website builders, browser agents and workflow managers. The resulting advice for founders is to avoid generic horizontal tools and aim at deeply specialized vertical niches, since foundation model providers appear likely to absorb the generic layer.
On cost, the assessment offered here is that ChatGPT delivers more tokens and operational headroom per dollar than Claude on both the $20 per month and $200 per month tiers, and that heavy agentic runs in tools like Claude Code or Cowork can reportedly use up a week's allotment in one prompt. Speed is described the same way: tasks that once took half an hour now complete in seconds. These are usage impressions rather than benchmark results, and the competitive lead is characterized as the current state of a race that keeps shifting.
Browser agents have also matured. Earlier tools depended on slow screenshots and text recognition and often refused to touch password fields. Current ChatGPT implementations can use saved credentials to sign in and navigate on the user's behalf, which is what makes end-to-end completion possible.
More cases: paper trading, media pipelines and photo libraries
A finance experiment connected ChatGPT to Alpaca Trader through its API. The prompt asked for a dividend portfolio with a $100,000 paper trading budget, $1,000 in paper buys per position and a forecast of future income. Paper trading means simulated orders with no real money. ChatGPT proposed ten stocks at $10,000 each, including Johnson & Johnson, Procter & Gamble, Coca-Cola and Realty Income, and projected about $3,691 in annual dividends, roughly a 3.69% yield or $308 per month. The Alpaca dashboard then showed the executed long positions and a balance of $99,969.89, down $30.11 on the day. These are simulation figures from one case; real trading carries volatility and risk, and the paper account is meant as a test ground before any real capital is involved.
Two mundane cases round out the picture. Riverside, a podcast recording platform, lacks bulk download and direct Google Drive sync, so ChatGPT Work was configured to log in, batch-download recording files and organize them into a shared Drive folder, a chore that human assistants had handled with delays and inconsistent naming. And for consumers, the same pattern could sort hundreds of vacation photos, pick the best and assemble a printable album, a task many people abandon for lack of time.
Key takeaways and what to do next
The pattern across every case is the same: give ChatGPT Work a goal and the integrations it needs, and it researches, builds, deploys and maintains the result in the background. If you want to try it:
- Create a dedicated Project for each recurring job so context, files and assets persist.
- State constraints up front, such as date, headcount, neighborhood, ages and preferences, so the browsing agent returns a precise shortlist.
- Ask for an internal Sites dashboard instead of a spreadsheet when a team needs shared visibility.
- Open long tasks with a goal specification and let Schedules run heavy work overnight.
- Consider an always-on desktop machine with a local MCP server if you want to direct builds from your phone.
- Keep financial experiments in paper trading accounts and treat any projected yield as a simulation.
For software companies, the consolidation of agent capabilities into ChatGPT itself suggests that durable products will come from deep domain specialization rather than general-purpose automation.


