ChatGPT Work: OpenAI ships the agent that finishes the job
OpenAI just moved the goalposts. On July 9 it launched ChatGPT Work, an agent that lives inside ChatGPT, takes an outcome instead of a prompt, and stays on a project for hours. You give it a goal, it pulls context across your apps, breaks the thing into steps, and comes back with finished materials β spreadsheets, slide decks, docs, even shareable web apps. Not a draft, not a chat reply. The deliverable.
What makes this more than a rebrand of the old Agent Mode is the engine underneath. Same day, OpenAI took the whole GPT-5.6 family to general availability: Sol the flagship, Terra the everyday workhorse, Luna the cheap-and-fast one. Sol sets a new state of the art on Terminal-Bench 2.1, the benchmark that actually tests planning and tool coordination in a shell, and posts big jumps on cybersecurity evals too. The preview shipped in late June under a US security review; the public version is here now, and ChatGPT Work runs on it.
The interesting part is the shape of the fight. OpenAI is no longer selling you a smarter chatbot β it is selling you a coworker that owns tasks end to end and reports back with artifacts. That is a direct shot at Microsoft Copilot, Salesforce, and every enterprise suite that bolted a chat box onto existing software. Rollout started with Pro, Enterprise, and Edu, expanding to Plus and Business over the following days.
The honest question is whether hours-long autonomy holds up outside a demo. Long-horizon agents still drift, and an agent that produces a polished-but-wrong deck is worse than one that asks. But the direction is unmistakable: the unit of AI work is shifting from the answer to the finished output. Details at https://openai.com/index/chatgpt-for-your-most-ambitious-work/ and the model at https://openai.com/index/gpt-5-6/
← Back to all articles
What makes this more than a rebrand of the old Agent Mode is the engine underneath. Same day, OpenAI took the whole GPT-5.6 family to general availability: Sol the flagship, Terra the everyday workhorse, Luna the cheap-and-fast one. Sol sets a new state of the art on Terminal-Bench 2.1, the benchmark that actually tests planning and tool coordination in a shell, and posts big jumps on cybersecurity evals too. The preview shipped in late June under a US security review; the public version is here now, and ChatGPT Work runs on it.
The interesting part is the shape of the fight. OpenAI is no longer selling you a smarter chatbot β it is selling you a coworker that owns tasks end to end and reports back with artifacts. That is a direct shot at Microsoft Copilot, Salesforce, and every enterprise suite that bolted a chat box onto existing software. Rollout started with Pro, Enterprise, and Edu, expanding to Plus and Business over the following days.
The honest question is whether hours-long autonomy holds up outside a demo. Long-horizon agents still drift, and an agent that produces a polished-but-wrong deck is worse than one that asks. But the direction is unmistakable: the unit of AI work is shifting from the answer to the finished output. Details at https://openai.com/index/chatgpt-for-your-most-ambitious-work/ and the model at https://openai.com/index/gpt-5-6/
Comments