OpenAI ships ChatGPT Work: an agent that writes your briefs, not your code

OpenAI launched ChatGPT Work today, a new agent mode that reads across a user’s tools, files, and browser to produce finished documents, spreadsheets, and presentations. Powered by GPT-5.6, the product is available immediately on macOS desktop, with Windows and mobile rollouts coming over the next few days. It works on all plans, including the free tier.

The product is not a coding assistant. It is a document-and-analysis agent. OpenAI is targeting the part of knowledge work that involves gathering context, synthesizing information, and producing formatted output. The company’s pitch: give ChatGPT Work a goal, let it pull from Slack, Google Sheets, Salesforce, Jira, and local files, then review and approve before it sends.

This is a bet that the highest-value near-term use of large language models is not writing code but writing the briefs, decks, and spreadsheets that occupy most office workers.

What ChatGPT Work actually does

The product is a new agent inside the existing ChatGPT interface. It operates on desktop where it can access local files, the browser, and approved applications. Users give it a natural-language goal, like “turn this product doc, launch timeline, and Slack thread into a launch brief with owners, risks, decisions, and next steps.” ChatGPT Work then assembles context, plans a sequence of actions, and produces the output.

Key design choices: the agent asks for permission before taking actions that send or share. It cites its sources. Users can schedule recurring tasks, like a weekly pipeline review that compares current quarter numbers against last week’s forecast and summarizes changes.

The use cases listed on the product page are revealing. They include weekly briefs, survey feedback analysis, email reply drafting, attendee list management, training follow-ups, and project tracker creation. None involve software engineering. The target user is not a developer but a program manager, marketing lead, or operations director.

OpenAI also published five customer testimonials. Nathan Bolt, Head of Digital Products at Virgin Atlantic, said ChatGPT Work turned a competitor analysis cycle “that would normally take weeks” into hours. Angela Ferrante at Zapier built a system that traced lead journeys across HubSpot, Gong, and email, generating visual maps and surfacing drop-offs. She credits the system with identifying seven figures in pipeline every month. Will Daney at NVIDIA said the tool cut his manual number-crunching time from 40% to zero. Chris Jones at Shopify uses it to run a research program across 3,500 non-R&D employees. Vaneet Seth at RingCentral scaled an early access program from 6 to 80 customers.

These are not AI researchers or software engineers. They are operators in large organizations.

The GPT-5.6 family

ChatGPT Work runs on GPT-5.6, which OpenAI describes as its “smartest model series for professional work.” The model comes in three variants: Sol, Terra, and Luna. The company says the model can “navigate ambiguity, adapt as work unfolds, and deliver polished outputs with less prompting.”

OpenAI has not published detailed benchmarks or a technical report for GPT-5.6 as of launch. The product page links to a research page that was not yet populated at the time of writing. The model’s capabilities are described in terms of workflow outcomes rather than standard evaluation metrics.

This is consistent with a broader industry shift. Labs are increasingly marketing models by what they enable users to do, not by how they score on MMLU or HumanEval. The product is the benchmark.

What this means for the AI economy

ChatGPT Work is a direct competitor to a growing category of AI-powered productivity tools. Microsoft’s Copilot, Google’s Gemini for Workspace, and a long tail of startups like Mem, Notion AI, and Granola all target the same problem: knowledge workers drowning in context switching and document production.

OpenAI’s advantage is distribution. ChatGPT Work works with any plan, including free. The company already has hundreds of millions of users. Adding an agent mode that reads across tools is a feature update, not a separate product purchase. Enterprises already on ChatGPT Business or Enterprise get it without a new procurement cycle.

The risk for OpenAI is that the product works well enough to change user behavior but not well enough to justify the compute cost. GPT-5.6 is likely expensive to run, especially when it needs to browse multiple tools and produce formatted output. OpenAI is betting that the retention and switching costs of an agent that knows your tools will justify the inference spend.

For the broader AI industry, ChatGPT Work validates a thesis that many have held but few have shipped at scale: the killer app for LLMs in the enterprise is not code generation but knowledge work automation. Code generation has a higher ceiling but a smaller addressable market. Every organization has people who write briefs and build decks. Not every organization has software engineers.

The open question

The product page lists what ChatGPT Work can do. It does not list what it cannot do. The testimonials describe successes. They do not describe failures, hallucinations, or the time the agent sent a draft with the wrong numbers.

OpenAI says the agent “asks for permission before taking action and cites its work.” That is the right design for a product that touches sensitive business data. But the safety and reliability of autonomous agents that read across multiple tools remains an open engineering problem. A hallucination in a code suggestion is a bug. A hallucination in a customer-facing brief is a reputation event.

ChatGPT Work is the most ambitious attempt yet to turn an LLM into a knowledge worker. The next few months will show whether the output is reliable enough to trust without checking every cell.