July 9 looked like one launch. It was actually two.
GPT-5.6 going public got the headlines — new models, new benchmarks, the whole ritual. But tucked into the same announcement was the thing that changes what ChatGPT is: a new mode called ChatGPT Work, where you stop chatting with the AI and start assigning it projects.
That distinction sounds subtle. It isn’t. Chat gives you answers. Work gives you deliverables — a finished spreadsheet, a formatted deck, a working draft — produced while you’re in a meeting doing something else.
Here’s the plain-English guide: what ChatGPT Work is, who gets it (including free users — genuinely), how to run your first task, and the fine print OpenAI’s launch page won’t lead with.
What is ChatGPT Work?
ChatGPT Work is an agent mode inside ChatGPT where you describe a project — “turn this CSV into a budget tracker with monthly summaries,” “build a deck from these meeting notes” — and ChatGPT goes off and does it, working in the background for minutes or even hours, then hands you the finished file. It launched July 9, 2026, alongside GPT-5.6, on the web and in a new desktop app.
OpenAI’s own framing, from the announcement: a partner for your “most ambitious work.” Strip the marketing and the honest version is: ChatGPT can now be delegated to, not just consulted.
If you’ve heard of Codex — OpenAI’s agent for programmers — ChatGPT Work is that same machinery pointed at office work instead of code. Spreadsheets, documents, decks, research briefs, even small web apps. It actually runs on Codex’s infrastructure under the hood, which matters later when we talk about usage limits.
Chat vs. Work — the mental model
The fastest way to understand the split:
The skill that matters changes with it. In chat, the craft is asking good questions. In Work, the craft is writing a good brief — the same skill you’d use delegating to a new hire. Vague brief in, confident-looking wrong thing out. That rule survives every model upgrade.
What it can actually do on day one
From OpenAI’s launch materials and what early users have been running this week:
- Spreadsheets with real structure. Paste messy data or connect a file; get back a workbook with tabs, formulas, and summaries — not a markdown table pretending to be one.
- Slide decks. Meeting notes or a rough outline in; a formatted, presentable deck out.
- Long documents. Reports, proposals, project plans — assembled and structured, not just drafted.
- Small web apps. A working signup form, a calculator, an internal dashboard. Nobody’s shipping production software from it, but “functional thing I can show my team” is honestly within reach.
- Multi-step research. Give it a question and sources; it reads, cross-references, and assembles a brief while you do literally anything else.
It plugs into your existing stuff through connectors — files, drives, the tools your ChatGPT account already links to. And on the new desktop app, Work sits one keystroke away all day, which is clearly the habit OpenAI wants to build.
Who gets it — including free users
This is the genuinely surprising part of the launch. The mapping, straight from OpenAI’s availability notes:
| Plan | ChatGPT Work access | Which GPT-5.6 model |
|---|---|---|
| Free / Go | ✓ Yes | Terra (fixed) |
| Plus | ✓ Yes | Sol, Terra or Luna — your pick |
| Pro / Business / Enterprise | ✓ Yes | Sol, Terra or Luna, plus the ultra setting |
Read that first row again. Free users don’t get GPT-5.6 in regular chat at all — but inside ChatGPT Work, they get Terra, a model that beats last generation’s flagship on several benchmarks. OpenAI put its newest tech behind the agent door, not the chat door. That tells you which behavior they’re trying to teach.
The dials, briefly: every plan with 5.6 access can toggle max (the “think longer” setting). The top setting, ultra — which runs four AI agents in parallel on your task — needs Pro or Enterprise inside Work. If you want the full Sol vs Terra vs Luna decision guide, we wrote that separately.
How to run your first Work task
- Open ChatGPT (web or desktop app) and switch to Work — it’s a workspace alongside your chats, not a buried setting.
- Connect what it needs. A file upload or a connector to where the real data lives. An agent without your context is just a chatbot with extra steps.
- Write the brief, not a prompt. State the deliverable (“an Excel budget tracker with a tab per month”), the audience, and what “done” looks like. Include one example if you have one.
- Pick the model if you’re on a paid plan. Terra for routine, Sol for anything with judgment in it. Skip ultra for your first runs.
- Let it run. Genuinely — close the tab. It works in the background and the desktop app pings you when it’s done.
- Review like a manager, not a fan. Open the file. Check the numbers against the source. Send back what’s wrong, specifically.
That last step is the whole game. Early users’ consistent finding this week: the output looks finished even when it’s 80% right. Polished-but-wrong is more dangerous than rough-but-wrong, because you’re tempted to skip the check.
The catch: it eats your usage allowance
Time for the part the launch page underplays.
Work tasks are expensive to run — a background agent thinks for minutes or hours, and every minute burns through your plan’s usage allowance. Because Work shares rate limits with Codex, one substantial task can take a visible bite out of your day’s capacity. Among early users, the loudest complaint of launch week wasn’t quality — it was watching a single few-minute task drain most of a usage window. And ultra multiplies that: four agents in parallel means roughly four agents’ worth of burn.
Practical translation:
- Treat Work tasks like meetings — schedule the ones that matter, don’t fire them off casually.
- Default to Terra, escalate to Sol only when the task has real judgment in it.
- Save ultra for the occasional genuinely hard thing, not your Tuesday reports.
- Batch your brief. One well-written brief that nails it in a single run costs a fraction of five sloppy attempts.
Does it actually work? The benchmark question
OpenAI’s launch numbers are strong — with the usual asterisks.
On Terminal-Bench, a hard benchmark of agent-style computer work, Sol scores 88.8%, rising to 91.9% with ultra’s four parallel agents. On GDPval-AA — a measure where human experts rate the quality of finished professional work, which is the closest benchmark to what ChatGPT Work actually does — Sol scores 1,747.8. Notably, Anthropic’s Claude Fable 5 scores 1,759.6 on that same table, published by OpenAI itself. On real-session coding evals run independently by Cursor (CursorBench), Claude’s best keeps a small quality lead while Sol wins clearly on cost.
Now the disclaimer, and it matters more than the scores. These are mostly vendor-published numbers from OpenAI’s own launch post — vendors pick which benchmarks to show and which settings to run them on. Benchmarks are also contest conditions: clean inputs, well-defined tasks, no ambiguity. Your job is none of those things. A model that wins a leaderboard can still fumble your expense report format, your client’s tone, your half-documented process. The only test that settles anything is running three of your real tasks through it and counting how much babysitting each one needed.
Which is honest advice in both directions — it might be worse than the chart suggests for your work, or better. Contest scores just can’t tell you.
What this means for you
If you’re on the free plan: you now have a real AI agent, today, at zero cost. Terra in ChatGPT Work is the best free AI work tool that exists right now. Start with one recurring task you hate — formatting a weekly report, cleaning a spreadsheet — and make Work do it.
If you’re on Plus: you got the biggest upgrade of anyone. Sol in chat, all three models in Work, for the same $20. The move is learning to split your work: chat for thinking, Work for producing.
If you manage a team: the skill to hire and train for is briefing — people who can specify a deliverable clearly get 10× more out of agents than people who prompt vaguely. That’s a writing skill, not a technical one.
If you’re wondering about Claude: yes, this is OpenAI’s answer to Anthropic’s Claude Cowork, which shipped its own big update two days earlier. The two take genuinely different approaches, and picking wrong costs you real money — we compared them head-to-head.
What ChatGPT Work won’t do
- It won’t know what “good” means at your company. It produces a competent generic version; the last 20% — your standards, your context — stays your job.
- It won’t check its own numbers. Verify anything that feeds a decision. Polished formatting is not accuracy.
- It won’t replace knowing the work. You can’t review a financial model you couldn’t have sketched yourself. Agents raise the value of judgment; they don’t retire it.
- It won’t stay still. This is week one. Features, limits, and model mappings will shift — expect the details above to age faster than the principles.
The bottom line
ChatGPT Work is the real story of the GPT-5.6 launch: OpenAI moving ChatGPT from “smart person you talk to” toward “assistant you delegate to,” and putting its newest models behind that door — including for free users. The tool is genuinely useful today, the usage costs are genuinely real, and the skill that decides whether it helps you is the oldest one in the book: saying clearly what you want.
That skill is learnable. Our ChatGPT Workspace Agents for Non-Engineers course teaches exactly this — briefing, reviewing, and building agent workflows without writing code — and AI Fundamentals covers the ground under it. Learn to delegate well and every agent, this one included, gets better for you.
Sources
- ChatGPT is now a partner for your most ambitious work — OpenAI (July 9, 2026)
- GPT-5.6: Frontier intelligence that scales with your ambition — OpenAI (July 9, 2026)
- OpenAI releases GPT-5.6 and ChatGPT Work tool — Axios
- OpenAI rolls out GPT-5.6 to the public on July 9 — Engadget
- CursorBench: model evals on real Cursor sessions — Cursor
