OpenAI has given ChatGPT a job description. ChatGPT Work, launched on 9 July 2026, is an agent that takes on whole pieces of office work: it gathers information across your connected apps, breaks a complex project into steps, and keeps going for hours until it hands back finished documents, spreadsheets, slides or even small web apps, as The Decoder reports.
The launch landed alongside the public rollout of GPT-5.6, and OpenAI paired it with a second announcement aimed squarely at the office: GPT-5.6 is now the preferred model inside Microsoft 365 Copilot. Between the two, the pitch is clear. The chatbot era is giving way to agents that do the work, not just describe it.
Update, 30 July 2026: OpenAI shows where the efficiency comes from
On 29 July OpenAI published an engineering breakdown of where GPT-5.6’s efficiency comes from, and it doubles as a status report on the agent story above. Rewritten GPU kernels cut end to end serving costs by 20%. Better speculative decoding lifted token generation efficiency by more than 15%. The agentic harness that Codex and ChatGPT Work share is a Rust layer that now caps any single tool’s output at 10,000 tokens.
The striking part is who did the work. OpenAI says GPT-5.6 Sol, running inside Codex, rewrote those production kernels itself, designed its own draft model across hundreds of experiments, and monitored the training runs, stepping in when hardware failed. A vendor pointing its own agent at its own inference stack is the clearest sign yet that this generation of tools has moved past demos.
One claim in the post does not survive a check. OpenAI says Sol beats Claude Fable 5 on Artificial Analysis’s Coding Agent Index at less than half the cost. On that index today Sol costs $7.08 per task against Fable 5’s $11.71, which is 39% less, not half. The same leaderboard now ranks Claude Code with Opus 5 first, ahead of Sol. We keep the tier by tier numbers in our guide to GPT-5.6 Sol, Terra and Luna.
What ChatGPT Work actually does
ChatGPT Work sits inside ChatGPT and behaves less like a chat window and more like a junior colleague. Give it a brief, and it plans the steps, pulls what it needs from connected business apps and workflows, and produces finished material: sheets, slides, documents and web apps. OpenAI says it can stay with a complex project for hours, completing steps independently rather than waiting for a prompt at every turn.
Under the hood it runs on GPT-5.6, the model family OpenAI released to the public this week in three sizes. If you want the numbers behind the new models, and how they compare with Anthropic’s, our GPT-5.6 vs Claude Fable 5 comparison covers what is verified and what is still vendor claimed.
Who gets ChatGPT Work first
Availability is staggered. ChatGPT Work is live for all users on the desktop app from day one, while the web and mobile versions roll out gradually by subscription tier. GPT-5.6 itself is reaching ChatGPT, Codex and the API over roughly 24 hours from launch. In practice, the fastest way to try the agent today is to install the desktop app and connect the tools your team already uses.
GPT-5.6 becomes Microsoft 365 Copilot’s preferred model
The same day, OpenAI announced that GPT-5.6 is now the preferred model in Microsoft 365 Copilot across Word, Excel and PowerPoint. The timing is pointed: the announcement arrived amid reports that Microsoft has been shifting parts of Copilot to its own in house MAI models, and TechCrunch notes that what preferred actually means is not entirely clear beyond OpenAI continuing to power Microsoft’s apps.
For businesses the practical read is simpler: if your team lives in Microsoft 365, the AI inside Word and Excel is being upgraded to the same generation of model that powers ChatGPT Work. Our ChatGPT vs Claude vs Copilot comparison breaks down how those plans stack up for Australian teams.

Why Australian businesses should pay attention
Agents that execute multi step work are exactly the category Australian businesses have been circling all year, and ChatGPT Work puts one in front of every desktop user overnight. The upside is real: repetitive reporting, document assembly and data pulls are the tasks teams most want off their plates. If the distinction between a chatbot and an agent still feels fuzzy, our guide to agentic AI vs generative AI is the primer.
The caution is governance. An agent is only useful once it can reach your apps, and that access is precisely what needs rules: which systems it may touch, what data leaves them, and who reviews the output. Launch week claims also deserve the usual scepticism until independent testing lands. Trial it on low risk workflows first, and measure before you migrate anything that matters.
The agent era is arriving on schedule
OpenAI turning ChatGPT into an agent, days after Anthropic pushed Claude’s scheduled cloud tasks, confirms where every vendor is heading in the second half of 2026. The tools will keep leapfrogging each other. What compounds for a business is picking a small number of workflows, wiring an agent to them properly, and letting the vendors fight over the model underneath.
