OpenAI spent three years teaching people to chat with an AI. On Thursday it started teaching them to hand work to one. ChatGPT Work, launched alongside the public release of the GPT-5.6 model family, doesn’t take a prompt so much as an outcome. It gathers what it needs from your connected apps and local files, breaks the job into steps, and works on its own — staying with a complex project for hours. What it returns isn’t a wall of chat. It’s the finished artifact: a spreadsheet, a slide deck, a document, or a working web app.
The delivery vehicle is its own land grab. ChatGPT Work runs inside a new ChatGPT desktop app for Mac and Windows that can reach across your local files, your installed apps, and a built-in browser. In the same stroke, OpenAI folded its Codex coding app into that desktop app, collapsing Chat, Work, and Codex into one “super app” available on every plan — including Free. Pro, Enterprise, and Edu customers can use ChatGPT Work today; Plus and Business get it over the next few days. Underneath it all sits GPT-5.6, whose cheapest tier, Luna, lists at $1 and $6 per million tokens — cheap enough to let an agent run for hours without the bill exploding.
That last detail is the whole strategy. An agent that thinks autonomously for hours only pencils out if the tokens it burns are cheap, which is why OpenAI shipped the price cuts and the agent in the same week. It also drops the company squarely into the fight it has circled all year — agentic knowledge work, the same ground Anthropic has been staking with its own document-, sheet-, and slide-producing agents. The pitch has quietly inverted. A year ago the product was a smart box you typed into. Now the product is the output, and the chat window is just where you place the order.
Our take: The thing everyone will screenshot is the finished slide deck. The number that actually matters is $1 per million input tokens. “Finished work, not chat” is only a business if an agent can grind for hours without every session turning into a four-figure invoice — so the real unlock here isn’t the agent, it’s the cheap tier underneath it. And watch what the app is allowed to touch: your files, your installed software, your browser. Whoever ships the deliverable also becomes the default place all the work lives.
What to watch
- The Plus and Business rollout. The mass market doesn’t sit on Pro and Enterprise. ChatGPT Work only matters at scale once it reaches the plans most people actually pay for — promised within “the next few days.”
- The token bill for long-running agents. An agent that runs for hours meters very differently than a chatbot you poke at. The finance teams that just started capping AI spend will feel autonomous work first.
- Who owns the desktop. Reaching into local files, installed apps, and a browser is the real moat. Every rival now needs an answer to “ships the finished file,” not just “answers the question.”
- Verification. An agent that hands back a polished deliverable also hands back polished mistakes — and the more finished the output looks, the less anyone checks it.