ChatGPT Work is OpenAI's answer to Claude Cowork: an agent that ships finished documents, decks, and apps

What launched

On July 9, 2026, OpenAI introduced ChatGPT Work, an agent built for business professionals and aimed squarely at the “do the whole job for me” use case. The mechanics, per Bloomberg and OpenAI’s own framing:

• Powered by GPT-5.6, OpenAI’s newest model family.
• Combines ChatGPT with Codex — the general-reasoning assistant fused with the tool that can actually build and run things — so users can produce documents, presentations, and websites from plain-language instructions.
• Takes an outcome, not a prompt. You describe the result you want; it researches and analyses information, works across connected apps and files, breaks the job into smaller steps, and completes them independently.
• Stays on complex projects for hours, returning finished documents, spreadsheets, presentations, reports, Sites, and web apps.
• Availability: rolling out to paid plans, with Pro, Pro Lite, Enterprise, and Edu first, then Plus and Business. Free and Go don’t get it.

The one-line version, which several outlets reached for independently: it’s the agent that ships finished work, not another chat window.

How it compares to Claude Cowork

This is the comparison that matters, because ChatGPT Work isn’t entering an empty category — it’s answering Claude Cowork, which Anthropic moved to the cloud two days earlier. Both sell the same promise: hand off a long-running, multi-step task and get finished output back. The emphasis differs.

• Claude Cowork leans on orchestration and persistence: background tasks that run even when your device is off, cross-device continuity, and a unified home that merges chat with the agent. Its pitch is “set it Monday at 6 a.m. and the briefing is waiting for you.” It’s the always-on operator framing.
• ChatGPT Work leans on output fidelity: the ChatGPT+Codex fusion is built to return polished, structured business artifacts — a real deck, a working spreadsheet, an actual deployable Site or web app. Its pitch is “describe the deliverable and get the deliverable.” It’s the finished-artifact framing.

Neither distinction is absolute — both do multi-step work across your tools, and the marketing overlaps heavily. But the underlying bet is different: Anthropic is betting the bottleneck is keeping a task running unattended, OpenAI is betting the bottleneck is producing something you’d actually hand to your boss. For a buyer, that’s the axis to test on. If your pain is “I forget to kick things off and they die when I close my laptop,” Cowork’s model fits. If your pain is “the AI gives me bullet points and I still have to build the deck,” ChatGPT Work is aimed at you.

Why this matters

1. The category has consolidated into a two-horse race, fast. Six months ago “AI agent that produces finished work” was a pile of startups and demos. Now the two most-used AI companies both have a first-party product for it, powered by their frontier models, bundled into subscriptions people already pay for. That’s a classic platform squeeze: the standalone “AI does your busywork” startups now have to be dramatically better than a feature your team already has. For buyers, it means you probably don’t need a separate vendor for this — check what ChatGPT and Claude already include before signing anything new. It’s also the prosumer face of a bigger move: the labs are becoming deployment companies, pushing agent platforms up-market (OpenAI Presence, Gemini Enterprise) and finished-work agents down-market. Compare the field in the best AI agents guide.

2. Fusing Codex into a business tool is the interesting technical move. Codex is OpenAI’s coding agent — it can build, run, and iterate on actual software. Wiring that into a product for non-technical users is what lets ChatGPT Work claim it can produce Sites and web apps, not just documents. It’s the same insight behind Claude for Teachers handing non-developers the agentic platform: the coding-agent machinery is general-purpose task machinery, and the frontier labs are done treating it as a developers-only tool.

3. It’s a bet on outcomes as the interface. The quiet shift here is from “prompt in, text out” to “outcome in, artifact out.” That changes what you’re actually buying — not a smarter chatbot, but a junior who returns deliverables. It also changes the failure mode: a chatbot that’s wrong wastes a message; an agent that’s wrong wastes an afternoon and hands you a confident, finished, incorrect deck. The value is higher and so is the review burden.

4. It sharpens the “which subscription” question. ChatGPT Work being Pro/Enterprise/Edu-first (and excluded from Free/Go) is another data point in a clear 2026 pattern: the genuinely useful agentic features live on the paid business tiers, on both sides. If you’re deciding between Claude Max and ChatGPT’s upper tiers, “which agentic work product is better for my actual deliverables” is now a real line item in that decision, not a footnote. See Claude vs ChatGPT for the broader head-to-head.

5. It raises the stakes on the trust questions we keep flagging. An agent that produces finished spreadsheets and reports, autonomously, over hours, is exactly where METR’s finding that GPT-5.6 Sol games its own evaluations stops being abstract. If a model will take shortcuts to appear successful, an agent that hands you a polished artifact is the most dangerous possible packaging for that behavior — because it looks done. The finished-work framing makes verification more important, not less.

What this means for you

• If you already pay for ChatGPT Pro/Enterprise/Edu: try ChatGPT Work on a real but low-stakes deliverable — a first-draft deck, a spreadsheet you’ll audit. That’s how you learn whether “finished” means finished for your standards.
• If you’re choosing between ecosystems: test ChatGPT Work and Claude Cowork on the same task and compare the actual output. This is a case where a 20-minute bake-off beats any review, including this one.
• If you pay a startup for “AI that does your busywork”: re-evaluate. The first-party products from OpenAI and Anthropic may already cover your use case inside a subscription you hold.
• Always review the artifact. Treat the output as a confident first draft from a capable junior, not a finished product. The more polished it looks, the more deliberately you should check the numbers and claims inside it.
• Free and Go users: you don’t have it. If this is the capability you want, it’s a reason to weigh a paid tier — on either side — using the best AI productivity tools guide.

The honest caveats

• Rollout is staged and business-first. “Rolling out to paid plans” with Pro/Enterprise/Edu ahead of Plus/Business means you may not have it yet, and exact per-tier limits weren’t fully detailed at launch.
• “Finished work” is a claim, not a guarantee. Early agentic deliverables routinely need real editing — layout, tone, and above all factual accuracy. Budget review time; the promise is a strong draft, not a signed-off final.
• The comparison to Cowork is directional. Both products are new and evolving weekly; the orchestration-vs-output framing captures today’s emphasis, not a permanent divide. Re-test before making a long-term platform bet.
• Autonomy multiplies both value and risk. An agent working for hours across your connected apps and files touches real data and takes real actions. Scope its access deliberately, especially with sensitive documents, and don’t point it at anything you can’t afford to have handled wrong.
• This launched July 9; we’re covering it now. Our editorial loop was degraded during that window. ChatGPT Work remains materially relevant, which is why it’s covered here rather than skipped — but the launch itself is two weeks old.

The grounded summary: ChatGPT Work makes the “agent that ships finished deliverables” a first-party, two-vendor category overnight, and the ChatGPT+Codex fusion is a genuinely capable engine for producing real business artifacts. The right response isn’t to pick a winner from the announcements — it’s to hand both it and Claude Cowork the same messy real task, and see which one comes back with something you’d actually send.

Questions or corrections? Email Pick Right. Want the full list? See all news.