The AI that finishes the work, not just the chat
OpenAI's new ChatGPT Work is an AI agent that takes a goal, plans the steps, pulls data from your tools, and hands you a finished deliverable — no more just asking questions.
- 3
- new AI model versions: Sol, Terra, and Luna
- $1–$5
- price to process one million input tokens
- 1,400+
- apps ChatGPT Work can connect to
- $30
- price per million output tokens from Sol, the top model
For two years, AI was something you talked to. You typed a question and got an answer. This week, OpenAI rolled out something different: ChatGPT Work. It is an agent — a program that takes a goal, makes a plan, pulls data from your tools, and comes back with a finished piece of work. Not a reply. A deliverable.
The practical side is straightforward. Describe what you need — compare your top five competitors and build a slide deck, for example. ChatGPT Work connects to more than 1,400 apps and services, gathers what it needs, and builds the output. It can run for hours without you watching. Set it to check in at each big step, and you stay in control and can catch mistakes early.
OpenAI also released three versions of its new model, GPT-5.6: Luna (fastest and cheapest), Terra (balanced), and Sol (most powerful). ChatGPT Work charges by how much the AI reads and writes, counted in tokens — small chunks of text. Luna costs $1 per million tokens in; Sol costs $5 in and $30 out. A large research task burns through a lot of tokens. Measure the cost before you automate something at scale.
Early testers report cutting multi-week analysis projects down to hours. That is real — but it is the optimistic case. Start with one task you already know how to check. AI agents make confident mistakes — they do not know what they do not know. Run it once with approvals turned on. If it passes that test, adding a second workflow gets much easier.